A field guide to the agentic era.

Tools & ReviewsPublished Oct 22, 202623 min read

llama-server Flags for Coding Agents: Context Size, GPU Layers, KV Cache, Slots and --jinja, Then Prove What Loaded

Tools & Reviews20 min read

CHOOSE YOUR NEXT MOVE01 — 03

Less searching. More understanding.

Get your bearings

What the new agent protocols, equal list prices and browser positions actually promise.

Build a better workflow

Steer or hand off, earn a schedule, test against a synthetic company.

Secure your fleet

Rotate keys before the cutoff, audit what agents publish, red-team every MCP integration.

THE INTEL LIBRARY

Follow your curiosity.

RSS feed

263 articlesGUIDES · REVIEWS · PERSPECTIVES

Claude Code Checkpoints and /rewind: Undo a Bad Edit, Know What Isn't Tracked, and Recover the Rest With Git

Claude Code checkpoints undo Claude's file edits, not Bash or manual changes. Run a /rewind drill, pick the right restore option, and recover the rest with git.

Tools & Reviews25 min read

Claude Code Subagents: Write .claude/agents Files, Restrict Their Tools, Pick the Model and Prove It Works

Create Claude Code subagents as .claude/agents files: three definitions, tool allowlists, model precedence, delegation tests, troubleshooting and rollback.

Tools & Reviews23 min read

llama-server Flags for Coding Agents: Context Size, GPU Layers, KV Cache, Slots and --jinja, Then Prove What Loaded

Set llama-server flags for a coding agent on purpose: context, GPU layers, KV cache, slots and --jinja, then prove them with /props, /slots and a tool call.

llama-server Remote Access for Coding Agents: SSH Tunnel, Tailscale Serve or Experimental --connect

llama-server remote access done safely: keep it on loopback behind an API key, then reach it via SSH tunnel, Tailscale Serve or the experimental --connect.

Tools & Reviews24 min read

Claude Code MCP Setup: Add Servers With claude mcp add, Pick the Right Scope, and Fix Connection Failures

Claude Code MCP setup, start to finish: add a stdio and an HTTP server, pick local, project or user scope, share .mcp.json safely, and fix failed connections.

Tools & Reviews22 min read

Claude Code Cron Job Setup: Choose /loop, Desktop, Cloud Tasks or claude -p, Then Add a Cost Gate

Set up a Claude Code cron job the right way: choose /loop, Desktop, cloud tasks or claude -p, then add a cheap pre-screen, spend cap, logs and a stop path.

Tools & Reviews21 min read

Claude Code VS Code Extension vs the Terminal CLI: Set It Up, Share Settings, and Fix Sessions That Won't Start

Install the Claude Code VS Code extension, decide when to use the panel or the terminal CLI, share one settings file, and fix sessions that won't start.

Tools & Reviews22 min read

Docker Sandboxes Claude Code Setup: Run sbx With a Chosen Workspace, a Tight Network Policy and a Clean Teardown

Docker Sandboxes Claude Code setup with sbx: pick a network preset, choose a bind, read-only or clone workspace, allow needed hosts, verify, then tear down.

Tools & Reviews22 min read

Migrate Branch Protection to Rulesets Without Opening a Gap for Your Coding Agents

Migrate branch protection to rulesets safely: snapshot classic rules, import in Evaluate or Disabled, map bypass lists, probe agent apps, then delete them.

Tools & Reviews30 min read

Claude Code Remote Control: Set It Up From Your Phone and Fix Disconnects

Set up Claude Code Remote Control in server mode under tmux, connect from your phone, read the saved spawn-mode messages and fix each documented error.

Tools & Reviews21 min read

Ollama Local Only Mode for Coding Agents: Enforce It, Then Prove It

Set up Ollama local only mode for Claude Code and OpenCode: find -cloud tags, set OLLAMA_NO_CLOUD on the real server process, and prove prompts stay local.

Git HTTPS Authentication GitHub Playbook: Authenticate Every Agent Clone Path Before It Stalls Your CI

Git HTTPS authentication GitHub guide: inventory each clone path, wire the right token per lane, and set a typed fallback before the next protection window.

Ollama q2_0 GGUF Import: Fit 2-Bit MoE Models, Do the Memory Math, Verify Quality

Ollama 0.40.6 reads q2_0 GGUFs. This ollama q2_0 gguf import runbook covers memory math, the Modelfile import, a q4 check, rollback and runner cleanup.

GitHub Copilot CLI Sandbox Proxy: Route Host-Rule Traffic Through HTTPS_PROXY After 1.0.100

GitHub Copilot CLI sandbox proxy runbook: route host-rule traffic through HTTPS_PROXY after 1.0.100, verify proxy logs, and keep NO_PROXY hosts direct.

Cline GPT-6 Configuration After 3.0.73: Fix max_completion_Tokens Lanes and Audit the New Defaults

Cline GPT-6 configuration broke on max_tokens. CLI 3.0.73 fixes it: restore GPT-6 lanes, set Azure OpenAI-compatible endpoints, and audit changed defaults.

DeepSeek Search API: Wire It Into Your Coding Agent Before Oct 31 Billing

Wire the DeepSeek Search API and Fetch API into a coding agent: token-bucket limits, authority and date filters, and a hard budget gate before Oct 31 billing.

Playbooks12 min read

Claude Code Fork Subagents: A Headless Migration Runbook

Claude Code fork subagents now reach headless and SDK lanes by default. Test inherited context, nested waits, permissions and failure evidence before promotion.

OpenCode Reasoning Variants: Validate Overrides After 1.19.3

Validate OpenCode reasoning variants after 1.19.3 with a lane inventory, configuration precedence checks, agent override tests and a clear rollback record.

Qwen Code Background Tasks: Require a Completion Receipt

Qwen Code background tasks need exit evidence, verified artifacts and resolved approvals. Apply a completion contract and five canaries before unattended runs.

llama.cpp Default Port Change: A v0.7.0 Upgrade Runbook

The llama.cpp default port change moves bare servers to 9931. Audit clients, preserve Docker mappings, test MoE cache budgets and coordinate RPC builds safely.

Claude Agent SDK Hook Rewrites: Prove the Input That Executes

Claude Agent SDK hook rewrites now undergo input validation. Use migration fixtures to verify executable arguments, preserve approvals and record outcomes.

LangGraph CLI Deployment: Promote the Tested Image to the Right Listener

LangGraph CLI deployment guide: promote a tested registry digest, choose listener placement at creation, and verify the resulting revision before rollout.

Tools & Reviews12 min read

OpenAI Codex: Startup and Configuration Diagnosis After 0.162.1

OpenAI Codex after 0.162.1: diagnose background-server startup, configuration layers, MCP failures and output contracts with a practical fleet diagnostic guide.

OSC 7501: The Agent Status Protocol Your Fleet Dashboard Should Consume

Claude Code 2.1.295 now speaks the OSC 7501 agent status protocol. Stop scraping spinner titles: consume the states, and fail closed when a guard hook breaks.

News & Trends17 min read

AI Agent Screenshot Leak: Audit Where Your Agents Publish Images After PixelLeak

PixelLeak found 13,000+ agent-posted images in public GitHub repos. Map where your agents publish screenshots, run a 30-minute audit, close the public paths.

Consistent AI Character Video Needs Approval Gates: The Viral Playbook, Rebuilt as an Agent Pipeline

A viral AI persona reached tens of millions of views with no label. Run character sheet, clip and schedule as gated stages with dated labels and a kill switch.

Industry Analysis17 min read

WebKit Says Agents Are Assistive Tech: The WebMCP vs ARIA Fight and Which Bet to Make

WebKit opposes WebMCP, Mozilla is neutral, one client ships. Use this surface-bet table to choose ARIA, WebMCP or both for each job your site lets agents do.

AI Moat After PhotoCraft: Clone-Test Your Own Product Before Agents Do

PhotoCraft rebuilt Photoshop's menus in a week; its README says behaviour lags far behind. Audit what a clone of your product could copy and what it would lack.

Verifiable Agent Consent: Keep a Delegated-Approval Receipt Ledger Before the First Dispute

Turn "the customer approved it" into evidence: the receipt ledger columns, what Decagon's PACT receipt carries and omits, and a timed drill for agent disputes.

OpenAI Legacy User API Keys Shut Off Oct 22: A Key-Migration Drill for Agent Fleets

OpenAI legacy user API keys are reported to stop working Oct 22. Inventory every key your agents read, rotate, canary-prove and revoke by Oct 20, with dates.

GPT-6 Interactive Answers Can Carry a Submit Button: Keep an Answer-Surface UI Register

ChatGPT's Intelligent UI puts forms and buttons inside answers. Classify every rendered control, gate the ones that act, log them, re-run when defaults flip.

Gemini 4 Argon Pricing Has Two Numbers: Budget the $4/$20 One in a Price-Expiry Register

Gemini 4 Argon launched at $2/$10 with no end date before $4/$20 applies. Book intro-priced models at list in a price-expiry register, then rehearse the step.

EmbeddingGemma 2 on a Laptop: Run a Twenty-Item Recall Test Before an Agent Searches Your Media

EmbeddingGemma 2 puts memos, screenshots and clips in one Apache 2.0 vector space. Prove recall on twenty of your own items and write an index manifest first.

Scheduled AI Agent Skills Earn Their Cron: The Manual-to-Skill-to-Cron Ladder Behind GBrain

GBrain's README claims 66 autonomous cron jobs. Here is the promotion checklist a skill must pass before it gets a schedule, plus a one-page overnight job spec.

AI Agent Red Teaming for MCP Integrations: Grade the Agent, Not Just the Model

Seven attack categories become seven test cases per MCP server, rerun on every version change and scored per agent. Anaconda's 300+ count is not the plan.

Kilo Desktop Puts 500+ Models Behind One Picker. The End of Lab Lock-In? Fill In a Mode-to-Model Table First

Kilo Desktop puts 500+ models, by Kilo's count, behind four agents. Before rollout, pin models per mode, log key owners and run a four-question buy test.

AI Agent Tool Gateway: Who Holds the Keys When One Token Calls Thousands of Paid Tools

One token can reach thousands of paid tool endpoints. Register who holds each key, run the --local check, cap per-agent spend and reconcile the meter monthly.

News & Trends16 min read

Personal Agent Protocol: Decide What Each Front Door Lets a Customer's Agent Do

Meta and Sierra announced the Personal Agent Protocol as a preview. Before v0.1 lands, register what each front door lets a customer's agent read and change.

The X Timeline Is Becoming a Task Queue. Here's Where a Tagged AI Bot Should Stop

Grok Bot's X connector reads your timeline; tag-to-Notion is unconfirmed. Sort nine verbs into allow, review and stop, and inventory every connected account.

Add a Critic, Not a Second Reviewer: Risk-Route AI Code Review Agents Like Meta's RADAR

Two AI reviewers scored worst in a Cornell/Stanford test; a critic of the review scored best. Route each PR by risk, as Meta's RADAR does, with this protocol.

GPT-6.1 Sol vs Claude Sonnet 5.5: Same List Price, Different Agent Bills. Price the Workload Shape

Both list at $2/$10 with $0.10 cache reads, yet a 50-turn agent loop past 272K costs 29% more on one. Price your workload shape with this five-row worksheet.

AI Agent Context Isolation: When a Finance Agent Shares a Computer With Your Slack Bot

A read-only agent on a shared computer can reach every sink its siblings have. Map sources against sinks, treat shared machines as one zone, gate posts as you.

A Reverse Engineering MCP Hit #1 on GitHub: Audit What Your Own Shipped App Reveals

An agent reverse engineering tool topped GitHub Trending. Audit what your shipped app reveals: keys, endpoints, symbols, update signing, plus a release gate.

Fast vs Slow AI Coding Model: Route Each Task to Live Steering or a Handoff

Fast serving tiers turned coding agents into something you steer mid-sentence. Route each task to steer or hand off, and meter tool time before buying speed.

Verify AI Agent Traffic in the Support Queue: A Triage Table for Customer Agents, Bots and Humans

Sort every chat, voice and API contact into verified agent, unverified agent, undeclared bot or human, with routes, limits, evidence and a platform kill switch.

Agents Need Fake Companies: Build a Simulated Company for AI Agent Testing

Era by Eon launched fake companies Oct 5; its benchmark grades by code, not AI judges. Write the fixture spec agents sit before a model swap: seed, graders.

WebMCP registerTool: Register Three Tools Before You Redesign Your Site

Make your site agent-callable without a redesign: register one read-only and two gated WebMCP tools with annotations, a kill path and an origin trial test.

Declined Means Stop: An AI Agent Consent State Machine and Fail-Closed Test Matrix

Treat every AI agent consent state except granted as a hard stop. One state machine for banners, GPC, MCP elicitation and OAuth, plus a release test matrix.

The AI Agent Pricing Page Is a Contract: Write Down What Is Never Paywalled

When agent credits run out, what still works? Put a never-paywalled table on your AI agent pricing page, and run the lapse test before you sign for one.

AEO for AI Agents: Get Classified as AI Before You Chase Citations

Directories and answer engines decide whether your agent product counts as AI before they cite it. Build the classification sheet that gets you past that gate.

AI Shopping Agent Checkout: The Pre-Authorization Checklist to Sign Before the Card Runs

Before an AI shopping agent checks out, sign seven lines: all-in total, caps, allowlist, expiry, scoped card, receipt and a revoke switch you have tested.

Claude Code Mods Are In-Process Code: Keep an Extension Manifest for Every Machine

Claude Code mods run in-process, can approve tool calls and aren't sandboxed. Keep a manifest per machine, pick a managed-settings policy, pin and roll back.

The Agent Stack14 min read

Claude Code OpenAI Compatible Proxy: The Bridge Register and the Removal Drill

Claude Code bridges expose an Anthropic-format endpoint and translate outward. Register each one, canary what it breaks, and pin allowedProviders every quarter.

Most Precise Tool First: A Desk-Job Inventory for Computer Use Agent Workflows

Half your desk jobs should never touch the mouse. Inventory each: precise path first, blast radius, kill switch and the evidence every computer-use run owes.

Meta Conversions API for Desktop App Install Attribution: A Four-Row Event Ledger

Meta's Conversions API has no slot for a desktop app install. Build a four-row event ledger with a consent gate, a 7-day send rule and a reconcile you own.

Docling vs MarkItDown: Run a Twenty-File Intake Bake-off on Your Own Documents

Run Docling vs MarkItDown on twenty of your own files: a per-file log, a hand-checked truth table, and a written rule for which converter gets each file class.

The Fast-Worker Lane Contract: Swap the Fast AI Coding Model, Keep the Lane

Fast AI coding models turn over in weeks. Write a lane contract for your cheap subagent worker, drill the swap on fixtures, and change one config line.

Review the Trajectory, Not the Reel: OSWorld 2.0's Checkpoints for Your Own Computer-Use Runs

Three OSWorld 2.0 numbers went around in September and no two measure the same thing. Borrow its checkpoint grader to review your own computer-use runs instead.

Own the Loop, Rent the Model: An AI Agent Harness Portability Register

Routers now rent shells, batches and key controls. Build a loop-ownership register, swap-test every part, and label what you rent before it quietly locks in.

One Login, Many Lanes: Keep Parallel Sessions Inside the Claude Code Usage Limit

Parallel Claude Code sessions share one usage window. Pin model and effort per lane, cap subagents, kill idle wakers and sum /usage across every machine.

Unofficial API or Official Connector: The Five-Rung Access Ladder for AI Agents After Reddit's API Shutdown

Reddit's API shutdown put dates on scraping lanes. Climb five rungs from official connector to wrapped session, and log every lane that risks an account.

The Standing Brief: Run a Read-Only Channel Agent on an Observability MCP Server

Anthropic gave a Slack channel a standing job and a Datadog MCP server. Start a rung lower: a read-only brief with owner, toolsets, evidence and a kill switch.

Tray Companion or Desktop AI Agent: The Decision Rule for Moving a Job

Which jobs should leave the tray for a desktop AI agent? A signal-by-signal decision table, what installs today, and a two-question exit test before any move.

Fetch Before You Drive: Web to Markdown for AI Agents on Research Jobs

Most agent research never needs a browser. Use this decision table to fetch pages as Markdown, log token headers, and escalate to computer use only when forced.

Memory Is an Input: Write-Provenance Rules for Shared Agent Memory

Shared agent memory lets one agent teach the rest of a repo. Inventory each store, record the source of every write and keep untrusted-input lanes read-only.

Cookie Custody: What a Viral Agent Tool Bag Asks You to Hand Over

Cookie custody for AI agent tools: register every session an agent holds, keep secondary accounts only, and ban install-by-prompt from remote instruction files.

One Run Contract Every CLI Loads: Stop Rules and Finish Lines in AGENTS.md

AGENTS.md stop rules only count if every CLI loads them. Write one run contract with a finish line, then prove Claude Code, Codex and Gemini CLI read it.

Banked Resets Are Perishable Capacity: Keep a Reset Ledger

A Claude rate limit reset refills one window, then expires. Keep a banked resets ledger: grant, expiry, owner, redeem surface, spend rule and a waste report.

The “Ask for More” Button Is an Intake Queue: Budget Requests With a Revert Date

Copilot budget increase requests are GA, and nothing in the docs reverts an approval. Run budget requests as a queue with required fields and a revert date.

Cost per Completed Task Beats $/MTok: Price the Finished Job, Not the Token

AI agent cost per task beats list price. Count attempts, retries, escalations, cache and tool meters, divide by tasks that finished, and compare models on that.

“Unconfigured” Is a Vendor Decision: Snapshot Every Agent Admin Toggle Before Oct 22

Copilot turns on Unconfigured features Oct 22 unless you decide first. Snapshot every agent admin toggle, diff it weekly and give each default an owner.

Kill Low-Effort Retry Loops: Price Effort per Completed Task

Reasoning effort cost looks cheap per attempt until you count retries. Record served effort, price effort per completed task, then pin a floor and ceiling.

Speed Is a Meter: Write a Speed Budget Before Anyone Types /fast

Claude fast mode cost is 2x standard for faster output only. Write a speed budget per lane: who may use /fast, opt-in, a monthly cap and tested kill switches.

“Free” Weights, Paid Tasks: Price Self-Hosted Overnight Runs per Completed Task

Self-hosted LLM cost per completed task: price hardware, power, operator minutes and failed overnight runs, then compare with hosted models at pennies a task.

Write a Finish Line the Transcript Can Prove

An agent goal completion condition is judged from the transcript alone. Write an end state, a stated check, constraints and a bound; back it with a Stop hook.

News & Trends17 min read

Leaked Model Names Are Not Candidates: A Rumor-Intake Rule

Leaked model names are not candidates. Admit a model only with an API ID, a price page, terms and a named provider, and park every rumor on a dated watch list.

The Evidence That Earns a Model the Default Slot

A vendor-set default model is a proposal. Pin the incumbent, then promote through shadow and canary lanes on merged-PR, revert and cost-per-task gates.

Never From a 6-Inch Screen: Write a Remote-Approval Policy

A remote approval policy for coding agents: what a phone may approve, what never goes remote, and a presence-marker hook plus a monthly drill to hold the line.

Review Bots Now Run PR Code: Where They Execute, What They See, What a Verdict Proves

Review bots now run PR code. Register where each one executes, which secrets and egress it gets and what its verdict proves, then drill it with a canary PR.

Runs Locally Is Not Stays Local: Audit What Your Agent's Remote Control Stores

Coding agent remote control keeps code on your desk but relays the transcript. Build a per-lane residency table, set off-switches, archive what stays local.

Who Actually Answered? Log the Served Model When Safety Fallbacks Swap It

Safety fallbacks can hand your request to an older model mid-run. Log the requested vs served model on every call, alert on mismatches, and fail closed.

TASKS.md With Receipts: Every Tick Needs Evidence

A TASKS.md checklist survives compaction, but a tick is only a claim. Give every tick a receipt a verifier can re-run, and revert any tick that has none.

Unofficial AI Wrappers: Your API Is Someone Else's Login, and the Account Is the Collateral

Unofficial AI wrappers turn a consumer login into an API, and a personal account becomes the collateral. Inventory each lane's credential and retire the risk.

Untrusted Telemetry In, Agent Code Out: Register Every Auto-Remediation Path

Auto-remediation lets error events, incidents and bot comments start coding agents. Register every path, gate the code-writing lanes, then drill with a canary.

Who Owns Compaction on an Unattended Lane: Notice, Warn, Force, and a Note-to-Self

Agent context compaction on unattended lanes: who fires it, token thresholds under price lines, what must survive the cut, and whether the summary is readable.

Jev Confidence Gates Before the Tool Runs

Set a Jev confidence threshold per tool class from 200 of your own labeled calls. Jev may deny or ask, never approve; destructive calls never run on its word.

GitLab MCP 19.4: Always-Ask Writes as the Default Dialect

GitLab MCP 19.4 turns fixed tool rules into settings: reads allow, writes ask, deletes deniable. Copy that dialect into every agent on your repos, then audit.

MCP Marketplace Security After MCPJacking and Plugin4Shell: Review, Pin, Verify

MCP marketplace security after MCPJacking's 155 hijackable entries and Plugin4Shell's SHA-pin bypass: an intake checklist to review, pin, verify and revoke.

Off-Peak AI Agent Scheduling: Route Flexible Work by the Clock, Not by Vibe

Off-peak AI agent scheduling by the clock: classify jobs, map DeepSeek's UTC peak hours and Anthropic's cache timers, route with caps, and review weekly.

OpenClaw vs AutoClaw: Buy the Metered Car or Build on the Engine

OpenClaw vs AutoClaw is an ownership question, not a price one. Decide who owns the gateway, skills, model and revoke path before a chat bot holds prod tokens.

Write Access Isn’t a Run Button: The GitHub Actions Trigger Allowlist for Agent Identities

GitHub Actions workflow execution protections decide which CI runs an agent can start. Strip edit and dispatch rights, map every cell, then drill each one.

Automatic Agent Session Cleanup: Audit the Settings Before They Fire

VS Code agent session cleanup, Devin archive and Codex delete can remove sessions, PRs and worktrees. Audit each setting, dry-run it and test restore first.

Stage-Only npm Tokens for Agents: What They Close and What They Don’t

An npm stage-only token blocks an agent's direct publish but can still move dist-tags and deprecate versions. Token classes, a drill, alerts, a 2FA review.

Canary Every CLI Upgrade: Catch Default and Compaction Drift on Unpinned Lanes

AI CLI upgrade testing for unattended lanes: record the model, effort, caps and compaction point a lane was served, replay every bump, fail undeclared drift.

Lint the Compaction Summary Before the Lane Acts on It

A compaction summary can tell the next context to hide mistakes. Lint it where each harness allows: before it loads, before the next tool call, or not at all.

Three Oversight Numbers Your Fleet Should Publish Internally

AI agent oversight metrics on Anthropic's own definitions: coverage before and after, review latency per leg, one escalation rate per monitor, every week.

Reconcile Before You Charge Back: Client Estimate, Gateway Meter, Provider Invoice

AI chargeback starts with reconciliation: price client estimate, gateway meter and provider invoice on one rate file, and block runs that don't tie out.

Prove Headless Runs Ignore State You Didn’t Declare

Claude Code bare mode plus a clean-versus-used differential run proves a headless lane loads no synced skills, AGENTS.md or memory its manifest didn't name.

Tools & Reviews17 min read

Routers You Can’t See Inside: Fugu Max, Copilot Auto, Cursor Auto vs Your Own Cascade

An LLM router that can't name the model behind each change fails change control. Five acceptance tests for Fugu Max, Copilot Auto, Cursor Auto and your cascade.

Use Jev to Route Cheap Models Without Guessing

Jev model routing as code: tiers with written criteria, an unclear exit, a floor that keeps unsure work on frontier, and cost per merged PR as the test.

Claude Code + Jev: Put the Gate in a Hook, Not an MCP Tool the Agent Can Skip

Claude Code Jev integrations shipped as a plugin and MCP tools the agent can skip. Put the gate in a PreToolUse hook that fails closed, with deny rules behind.

Log Every Jev Decision — and Pin the Version the Record Carries

A Jev decision log needs one record per call: state reference, versioned questions, answers, reported model and policy. Pin jev-1.13.0, never jev-latest.

Ship Typed Jev Decisions on Vercel, Cloudflare or OpenRouter — Test the Dialect Before You Pick the Door

Vercel AI Gateway's evaluation API calls Jev's Noul a boolean. Run one fixture set through Vercel, Cloudflare, OpenRouter and TypeSafe, then pin one wire shape.

When Jev Returns 429: Decide Fail-Open or Fail-Closed First

A Jev rate limit will stall your gates. Write each gate's fail mode first, budget both TypeSafe limits fleet-wide, and cap SDK retries inside the deadline.

Jev Can Be Prompt-Injected Too: Read the Jaggedness Page Before It Guards a Tool

Jev prompt injection is in TypeSafe's own docs. Put an allowlist first, send only needed fields, and pin injected fixtures so a flipped verdict blocks upgrades.

Fan Out Jev Judgments, Compose Policy in Code

Jev API questions belong in one call: keyed Nouls, Choices and Scores return a decision vector. Compose policy in code, validate IDs, retire yes/no subagents.

Shadow-Mode Jev Before It Skips Your CI Tests

AI CI test selection with Jev starts in shadow: count false skips against real failures, canary on a slice, and pin the action by SHA before you enforce.

Jev Soft-Ranks Skills — Agents Still Decide

Jev skill selection without hiding skills: gate on need, shortlist three, inject one suggestion line, let the agent decide, and log every override it makes.

Industry Analysis17 min read

Jev Isn’t a Chatbot — Stop Routing Like One

TypeSafe Jev belongs in a decision seat, not a reply seat. Sort fleet decisions with four tests, keep judges and compactors off it, and ship the table.

News & Trends18 min read

LiteLLM MCP on the KEV List: A Patch Clock for Agent Gateways

The LiteLLM MCP vulnerability hit CISA KEV on Sep 2. Block /mcp/ at the proxy, start the clock at the advisory, and prove a forged Bearer now gets a 401.

Managed Permissions Users Can’t Override — Test It Across Harnesses

Copilot enterprise managed permissions claim users can't override them. Prove it: one deny, ten bypass rows, run in Copilot, Claude Code, Codex and JetBrains.

The Rule of Two: Split the Lane That Holds All Three

The agents Rule of Two for a real fleet: score each lane on untrusted input, private data and egress, split all-three lanes, and lint configs before they run.

Test What Survives a Cross-CLI Session Import

Import Claude Code session history into another tool only after a fidelity test: mark turns, tool calls, approvals, branch and model carried or dropped.

Playbooks20 min read

A Renamed Tool Is a Disabled Guard: Contract-Test Hook Coverage Before Antigravity’s Oct 5 Cutover

Antigravity agent hooks match tool names and fail open, and 09-2026 renamed the file tools. Test guard coverage and replay wire fixtures before Oct 5.

Unauthored Hooks Are Persistence: Remove Them and Prove They Stay Gone

Claude Code hooks malware survives a declined consent prompt. Baseline hooks, logon tasks and PATH shims off-box, allowlist helpers, prove removal holds.

Your Vendor Lane May Not Be the Last Hop

AI vendor data routing can carry your prompt past the vendor you signed with. Register each lane's last hop, then cap its data class at what you can verify.

Vendor Permission Verdicts: Which Ones Still Need Your Approval?

Claude Managed Agents permission policy verdicts, normalized: log allows and denials as evidence, queue only real asks, and mark unseen verdicts unavailable.

One Checkout, One Owner: Native CLI Worktrees vs Your Lane Runner

Codex, Claude Code and Cursor now create git worktrees for AI agents. Give each checkout one owner, lock what your runner owns and prove it with a decoy.

Annotation Is Not Approval: The Human in the Loop Annotation Approval Split

Human in the loop annotation approval blurs when note-taking sits beside deploy, pin and schedule buttons. Tier gestures by consequence; check the gate weekly.

AutoClaw Connectors: The Weekly Blast-Radius Inventory

AutoClaw connectors ship with no scope or revoke docs. Run a weekly inventory: what is installed, who added it, last used, write scope and a tested revoke path.

AutoClaw: Inventory the Digital Employee Before Cluster Mode Fans Out

AutoClaw is a digital employee on the machine. Inventory credit plan, bot tokens, connectors, Cluster Mode, wake-lock and Hermes gate, then test the kill path.

AWS MCP Server Lambda Triage: Split Diagnostic Reads From Deploy Writes

The AWS MCP Server Lambda capability triages in one read-only call. Keep deploy writes on a separate IAM role, MCP config entry and human gate, logged apart.

Agent Dashboard Cards: One Operating Metaphor per Surface, or Your Team Drowns in Toast

Agent dashboard cards now mean four things: widgets, sub-agents, deliverables, threads. A dialect map of what each card is, what it can do, who owns the kill.

News & Trends17 min read

Claude Code Projects: Anthropic's Cloud Coordinator, and Who Owns Abort When Threads Spawn

Claude Code Projects runs cloud threads on their own branches under one coordinator. Stop, Pause, Archive, Delete semantics, 200 threads a day, who owns abort.

AutoClaw Cluster Mode Progress Panels vs Tray Stall Flags: Two Ways to See an Agent Is Still Working

AutoClaw Cluster Mode shows step status in chat; a tray shows stall flags. Two ways to see still working, neither a kill switch. Who owns abort on one laptop.

Computer Use Agent Windows Policy: Entitlements Before the Mouse Moves

A computer use agent windows policy for corporate endpoints: approved models, standard-user driver accounts, an entitlement matrix, a kill switch and logging.

AI Credits vs Tokens: Two Billing Dialects, One Operator Ledger

AI credits vs tokens are two billing dialects. Translate a credit pack into measured jobs, price the same jobs at token rates, and keep one burn-rate ledger.

Cyber Capability Gates Are Becoming Product SKUs: Allowlist AI Cyber Model Access by Agent Role

Gemini 3.8 Flash Cyber, Mythos 5.1 and Astra Daybreak turned cyber capability into gated SKUs. Map AI cyber model access by agent role, log it, rehearse revoke.

DeepSeek V4.1-Flash: Rewrite the Off-Peak and Cache-Hit Runbook

DeepSeek V4.1-Flash halves prices off-peak and bills cache hits at $0.006/MTok. The UTC windows, the prompt shape that hits cache, and the re-based budgets.

Claude Fable 5.1 Cache Reads: Meter Dollars per Overnight Job

Claude Fable 5.1 cache reads cost 0.025x base input. Meter dollars per overnight agent job and cache-hit rate, learn what breaks the cache, set effort per job.

Model Deprecation Routing Is a Continuity Drill, Not a Changelog Line

Model deprecation routing runbook: when a provider swaps the model under your ID, run ordered fallbacks, smoke tests, spend caps per lane and human notify.

GitLab Duo Triggers on Merge Request Created, and the Loops Around Them

GitLab Duo triggers fire on merge request created. Bots cannot activate them, so the loops left are external agents, cross-system hops and cost. Five guards.

GPT-6 Astra Computer Use: Inventory Every Host That Can Drive the Mouse

GPT-6 Astra computer use makes every screen a model can reach a trust tier. Build the host ledger first: who can drive it, which account, kill switch, evidence.

IM Channel Allowlists for Digital Employees: Gate the AutoClaw Telegram WhatsApp Bot Before It Joins #ops

An AutoClaw Telegram WhatsApp bot can drive a local agent from any group it joins. Write the channel allowlist, rate cap and revoke path before it hits #ops.

A Keyless MCP Server Is Its Own Trust Tier

A keyless MCP server has no grant to revoke, no published rate limit and feeds open-world text into context. The allowlist tier, five controls, a probe script.

Kimi Work Dashboard Cards vs Automater Lite Home Cards: The Widget and the Status Wall

Kimi Work Dashboard cards are widgets the agent builds; Automater Lite Home cards are a status wall for running CLIs. The decision table, limits and kill rules.

Local AI Agent vs Cloud Project: Digital Employees on the Desk, Threads in the Cloud, and No Sync Fantasy

Local AI agent vs cloud project: AutoClaw, OpenClaw and Kimi Work beside Cursor and Claude Code Projects; one identity, one kill switch, provenance per host.

Misalignment Disclosures Are a Fleet Evidence Problem: AI Agent Incident Evidence Before the Session Rotates

OpenAI disclosed six misalignment classes from its own runs. Build AI agent incident evidence for your fleet: tool trail, approvals, cost, diff and environment.

Kimi Pin to Desktop Allowlist: Always-Visible Agents Need a Gate

Kimi Pin to Desktop makes a widget always-visible. Write a pin allowlist like an MCP allowlist: who may pin, what data it may show, always-on-top and expiry.

Plugin4Shell: SHA Pinning Without HEAD Verify Is Theater

Plugin4Shell beat SHA pinning in four coding agents. A Tuesday drill: version floors by harness, post-checkout HEAD assert, auto-update posture, plugin census.

Three Meters for One Agent Job: Inventory Your AI Agent Billing Meters Before Fan-Out

One agent job can hit three ai agent billing meters: tokens with cache multipliers, per-call tools, and sandbox minutes with a 5-minute minimum. Cap each one.

Multi-Agent Abort With Three Coordinators, One Human, and No Shared Bus

Multi-agent abort runbook for Claude Code Projects, Cursor Projects and the OpenAI Agents API: one human interrupt path, evidence export per vendor, pause all.

Kimi Work Scheduled Tasks Are Unattended Desktop Cron: Tier Them Like a Headless Run

Kimi Work scheduled tasks turn a live widget into unattended desktop cron. Tier it like a headless run: allowlists, no prod creds, run records, a human gate.

Zero Data Retention AI Agents: Where Vendor Forgetting Meets Agent Memory

Zero data retention AI agents need memory to finish overnight jobs. What Anthropic's Covered Models and OpenAI's endpoint table retain, plus a boundary table.

AI Agent Cost Alerts: Anomaly Thresholds, Paging, and Auto-Pause for Agent Fleets

AI agent cost alerts for fleets: baseline two weeks, define a 3x-day anomaly plus velocity and worker triggers, page with the right facts, pause spawns first.

AI Agent Environment Setup Is Two-Thirds of the Failure: Fleet Readiness Before the Prompt

AI agent environment setup caused 65% of GitTaskBench failures. Run this six-gate preflight (image, lockfile, toolchain, smoke test, budget, score) first.

Agent Evaluation CI Gates: Fail the PR When the Agent Regresses

Build an agent evaluation CI gate: deploy the agent, run a fixed prompt set, score its tool choices, and block the PR on regression. Thresholds and YAML inside.

AI Agent Identity as a Service Principal: Stop Sharing Human SSO

Give every AI agent identity of its own: a service principal, GitHub App, or IAM role, scoped grants, short tokens, a broker, and revocation that spares users.

CI Failure to Fix: The AI Agent CI Loop That Doesn't Thrash

An AI agent CI loop that retries every red check will thrash. Cap attempts at three, gate on new failure signatures, set a cost ceiling per PR, then hand off.

Tools & Reviews13 min read

Nightfall's MCP Gateway: Buy the Proxy or Build the Six Checks

A buying guide for the MCP gateway decision: score Nightfall's proxy against six checks, run a two-week acceptance test, and see what no SaaS gateway covers.

Cursor Projects vs. the Tray: Who Owns Your Fleet Coordinator

Cursor Projects puts a fleet coordinator in the IDE. The ownership table: what it runs, what a local tray owns (stall flags, kill switch), and what breaks.

GitSpawn Week: Untrusted Repos Are an Intake Problem

GitSpawn turns opening a folder into code execution. A Tuesday intake checklist: clone-only policy, read .git/config first, version floors, a quarantine user.

codex exec Headless Modes: Unattended Is a Different Trust Tier

codex exec headless runs, Claude Code print mode, and Actions wrappers need their own trust tier: flag shapes per tier, no prod creds, an evidence pack per run.

Human in the Loop Approval Fatigue: Queues That Don't Become Rubber Stamps

Human in the loop approval fatigue turns agent approvals into rubber stamps. Tier by consequence, batch the low-risk, expire stale prompts, measure the reflex.

Hybrid Agent Fleets: Laptop + Cloud VM Without the Sync-Everything Fantasy

Run a hybrid agent fleet across laptop, cloud VM, and a vendor's computer: per-host identity, transcript provenance, a kill switch per host, git-only hand-offs.

Interrupt AI Agent Coordinators Safely: What Pause, Redirect, and Abort Must Mean

How to interrupt AI agent coordinators without orphaning work: a signal ladder, tool-boundary stops, branch-per-worker git rules, a redirect protocol, a drill.

Managed Agents Comparison: Bedrock, Claude, OpenAI. Pick Failover First.

A managed agents comparison of Bedrock AgentCore, Claude Managed Agents, and the OpenAI Agents API on state, tool schemas, evidence, kill switch, and residency.

MCP Server Inventory Before Allowlist: The Weekly Ritual

Run a 45-minute weekly MCP server inventory: find every config, attribute each server, score blast radius, diff versions, then keep, prune, or pin each row.

The Agent Memory Benchmark You Can Run on a Tuesday

Run an agent memory benchmark on your repos in one afternoon: a frozen suite, three arms, P99 latency, cost per 1k lookups, harm cases, and a decision rule.

OpenAI Agents API Is a Managed Harness. Run the Continuity Drill Anyway.

The OpenAI Agents API rents you the Codex loop. Inventory model, harness, and sandbox dependencies, write failover routes, and keep the record on your disk.

Overnight Agents: The AI Agent Merge Gates That Define Done

Overnight agents open PRs while you sleep. AI agent merge gates define done: tests, a diff ceiling, secret scan, path rules, an eval threshold, human approval.

AI PR Review Agent Policy: Agents Propose, Humans Merge

An AI PR review agent should propose, never merge. The policy: always-human paths, a CODEOWNERS shape, no bot-approves-bot, and reviewer quality you can track.

Industry Analysis13 min read

From Dashboard Theater to Pre-Action Gates: Where AI Agent Security Gates Belong

AI agent security gates should act before each tool call: approve, prune, broker, revoke, and record. Use the matrix, vendor questions, and Tuesday drill.

News & Trends13 min read

The Sandbox Is a Suggestion: What Black Hat's AI Agent Sandbox Escapes Broke for Operators

An AI agent sandbox escape hit Claude Code Action, Gemini CLI, and Codex at Black Hat 2026. The compensating controls operators can install this week.

Slack AI Agent Subscriptions Without a Control Plane

Slack AI agent subscriptions turn every channel message into a worker. The runbook: channel allowlist, spawn cap, event dedupe, human gate on merge, revoke path

When the Coordinator Spawns a Thousand Subagents: Metering Subagent Token Cost at Fan-Out

Subagent token cost climbs one worker at a time. Cap concurrent workers, budget per task, log every spawn, and kill orphans before a coordinator fans out.

Claude Code Permission Modes, Codex Sandboxes, One House Dialect

Map Claude Code permission modes and Codex sandbox flags onto three house tiers, encode them once, audit which host drifted, and learn what subagents inherit.

Evidence When the Harness Is Vendor-Hosted: Export the AI Agent Audit Trail Before Access Changes

Vendors keep the agent session record. Export the AI agent audit trail: tool calls, approvals, costs, final diff, and environment before access changes.

Industry Analysis12 min read

When Not to Use an AI Agent Coordinator

An AI agent coordinator copies its first mistake N times. The decision rule, a sixty-second self-test, and four brakes for sensitive or air-gapped work.

Playbooks12 min read

Windows IT After GitSpawn: An AI Coding Agent Endpoint Policy

After GitSpawn, an AI coding agent endpoint policy for Windows IT: standard-user accounts, AppLocker version floors, git overrides by policy, intake quarantine.

News & Trends11 min read

Rehearse the Provider Cutoff: Model Failover for Agent Fleets

Use the proposed Cursor cutoff to rehearse model provider failover: inventory dependencies, validate supported routes, test quality, and price capacity.

Playbooks11 min read

Open-Source Agent Gateway vs Vendor Security Suite: Pick the Layer You Refuse to Lock In

Agent gateway open source vs vendor suite: a six-layer scoring runbook for which control-plane layers stay portable, plus an exit test and vendor questions.

Playbooks11 min read

AI Agent Identity: Treat Every Agent as a Privileged User

AI agent identity runbook: workload identities, scoped grants, credential brokers, provider expiry limits, and revocation tests across each trust boundary.

News & Trends10 min read

GPT-6 Astra at the Critical Cyber Threshold: An Operator’s Runbook

Assess GPT-6 Astra enterprise access and prepare production controls: contain active agents, reconstruct their actions, and gate consequential writes.

The Agent Gateway Is the Control Plane for Enterprise Agents

An agent gateway is the control plane between enterprise agents and their tools. Run six checks this week: access, approvals, secrets, audit, revoke, tenancy.

Deadbugz Killed Approve-Once: The Runtime Controls MCP Needs Now

Deadbugz hid malicious MCP metadata behind ordinary calls. Build a runtime loop with pinned manifests, re-approval, bounded egress, and call-time evidence.

Playbooks11 min read

Shadow MCP Is the New Shadow IT

Shadow MCP is the new shadow IT. One-week runbook: sweep harness configs, build an approved MCP inventory, quarantine unregistered servers, catch drift nightly.

News & Trends10 min read

Stateless MCP Is a Gift to Gateways

Route stateless MCP by validated headers, separate transport logs from tool outcomes, and migrate legacy clients with explicit policies, cache keys, and checks.

Product9 min read

A Day on the Windows Tray

A composite day of Windows AI fleet management: recover after a reboot, inspect a stalled session, review a diff, meter parallel work, and redact a secret.

Cold Restarts on Windows: Sessions That Don't Come Back

Recover AI sessions after a Windows reboot: verify transcripts, restore WSL and Docker bottom-up, inspect the working tree, then resume or re-brief each agent.

Agentic Foundations8 min read

Companion vs Harness vs Computer

AI companion vs harness vs computer-use agent vs ADE: define each layer, map who commands whom, and identify the capability a product actually sells.

You Can't Replay What You Can't See

An incident playbook for AI session replay: search supported session records together, inspect tool calls, match them to the diff, and resume where supported.

Memory That Burns Quota

Measure Claude memory cost across auto memory, CLAUDE.md, and plugins, then replace indiscriminate replay with a bounded, archive-first retrieval policy.

Claude Code Permission Modes Are Fleet Policy

Choose Claude Code permission modes by repository trust, define who can escalate them, and audit the same policy across an AI-agent fleet.

Stop Trusting the Mermaid

Review AI-generated diagrams as claims: inspect the transcript, diff, command outputs and current tests, then reproduce suspicious behavior before merge.

Plugins, Gists, and Rotting AGENTS.md

AGENTS.md files drift, conflict, and multiply until no two agents run the same job. The playbook: what stays, what becomes a skill, and the quarterly rot audit.

Agentic Foundations7 min read

'AI With No Restrictions' Usually Means Local

Most people searching for AI with no restrictions don't want jailbreaks. They want local: models, transcripts, and installs no vendor can cap or cut off.

Muse Code's Session Bus: Inter-Session Messaging as the Operator Primitive

The Muse Code session bus is live: inter-session messaging over a local socket, plans from $5. What the primitive does to visibility, token burn, and replay.

Tools & Reviews11 min read

Perplexity Personal Computer vs a Windows Tray Companion

Compare Perplexity Personal Computer for Windows with a local tray companion across role, price, data boundary, platform, and vendor risk.

When the Harness Is a Plugin Bag, You Still Need a Boss

DeepSeek Harness makes the loop, sandbox, and model swappable plugins. That eases agent harness lock-in — and still leaves your fleet without a boss layer.

Agentic Foundations7 min read

What Is an AI Computer? A Daily-Driver Definition

What is an AI computer? One name covers three layers — the machine, the computer-use agent, and the operating layer. A definition with 2026's products mapped.

Claude Code After September 14: The 17% Cut You Will Feel

Claude Code limits change September 14: the +50% boost ends and settles at +25%, a 17% reduction from the temporary allowance. See the math and checklist.

Cross-CLI Memory That Survives the Session

AI session memory that outlives one tool: keep supported histories searchable and local, keep preferences small, and keep secrets out of both.

Cursor After SpaceX: The Proposed Cutoff Is a Command-Center Problem

SpaceX closed its Cursor acquisition August 14; OpenAI proposed ending model access November 12. Here is the record and a practical continuity checklist.

Product10 min read

The Desktop ADE: Session Explorer, WSL, Docker, and the Topology Chat Never Shows

Tour the Automater Desktop beta: searchable cross-provider Session Explorer plus a separate live topology for WSL, Docker stacks, containers, and hosts.

Product9 min read

Many Assistants, One Boss — What Grok Bot Changes

What Grok Bot is, where its data lives, and how to manage it beside local AI CLIs without blurring cloud storage, permissions, or transcript boundaries.

Product10 min read

Stall Flags and Keepalive: AgentOps From the System Tray

AI agent monitoring from the Windows tray: what a stall flag means, amber vs. green, what keepalive prevents, and what a tray honestly can't fix.

Agentic Foundations8 min read

What People Mean by 'AI Computer' on Windows in 2026

Searching for AI computers? On Windows in 2026 the phrase means agents on the PC you own — computer-use workers, the tray boss that runs them, no new hardware.

Product9 min read

Automater Lite: The Windows Tray That Watches Your AI Fleet

Automater Lite is a free Windows tray companion with a local AI-session Library, fleet status, search, usage meters, and signed updates. See where PRO begins.

News & Trends10 min read

Corporate AI on a Windows PC: What IT Is Actually Buying

IT is being asked to put AI agents on company computers. A corporate AI checklist that works: where sessions live, what leaves disk, who sees the fleet.

Tools & Reviews8 min read

Home AI Without a Dedicated AI Workstation

Build a useful home AI setup on the PC you own. Learn when local inference hardware earns its cost, and when archive, monitoring, and metering matter more.

Local-First AI: The Vault and Session Archives That Stay Local

Local-first AI as operating practice: keep the session archive on your disk, scrub secrets before indexing, and map every optional connected data path.

Tools & Reviews9 min read

The Operating Bill, Not the Token Bill: Price the Column Your AI Budget Skips

Your AI subscription cost is only half the bill. Price the other half, the operating bill, against the hidden bill of lost sessions, with two fleet scenarios.

EU AI Act Article 50 Is Live: What Agent Builders Must Disclose Now

EU AI Act Article 50 took effect August 2, 2026. What agent builders must disclose, how to mark AI content, who is in scope, and a practical checklist.

Evals for AI Agents: How to Measure Software That Acts

What are evals in AI? A plain definition, four grader types, pass@k worked examples, and a five-step plan for your first agent eval suite in one week.

Agentic Foundations13 min read

Types of AI Agents, Explained With Real 2026 Examples

The five textbook types of agents in AI, then the 2026 taxonomy that matters: four axes, eight real tools mapped, and a straight answer about Copilot.

Agentic Foundations14 min read

Agentic AI vs Generative AI: What Actually Changes When Software Gets Goals

Agentic AI vs generative AI, minus the vendor gloss: the architecture gap, a real comparison table, cost and risk asymmetries, and when a plain prompt wins.

LangGraph in 2026: Deep Review, Real Use Cases, and When to Use Something Else

An honest LangGraph review for 2026: the graph model, checkpointing and interrupts, three real builds, platform pricing scrutiny, and when to skip it.

Beyond the Terminal: Cowork Tools and Autonomous AI Agent Workspaces

What an AI agent workspace really delivers in 2026 — Manus, Genspark, Devin, ChatGPT agent, and Claude Cowork compared on task fit, pricing, and trust.

MCP Goes Stateless: What the 2026-07-28 Spec Changes for Builders

The 2026-07-28 MCP spec retires sessions, replaces elicitation with MRTR, and hardens OAuth. What changed, why, and how to migrate servers and clients.

Amp, Crush, and OpenClaw: The New-Wave Harnesses Worth Watching

Eight new AI coding tools mapped: Amp, Crush, OpenClaw, Ecodex, and more — with evidence tiers, survivor criteria, and a safe two-week trial protocol for 2026.

Agentic Foundations15 min read

AgentOps: Running a Fleet of AI Agents Without Losing Your Mind

Agentic ops, defined by people who run agent fleets daily: the five-layer AgentOps stack, the four metrics that matter, and a starter incident runbook.

DeepSeek Harness: The "Everything Is a Plugin" Agent Runtime, Reviewed

DeepSeek Harness reviewed: the MIT agent runtime where everything is a plugin. Four modes, a session-log core, real sandboxing — and who should switch now.

GitHub Copilot CLI, Grok, and Amazon Q: The Other American Harnesses

GitHub Copilot CLI after the June 2026 AI-credit switch, Grok's missing CLI, Amazon Q's blocked signups: one honest review of the second-tier US harnesses.

The Agent Stack17 min read

Securing AI Agents: Prompt Injection, Least Privilege, and the New Attack Surface

AI agent security in practice: the lethal trifecta, prompt injection, MCP hardening per the June 2026 government guidance, and controls that bound blast radius.

Databricks and the Enterprise Agent Platform Race

What Databricks is, who owns it, and whether the lakehouse can own enterprise agents: the data-gravity thesis, the MCP counter-case, and a verdict by workload.

The Great Harness Die-Off: Gemini CLI, iFlow, and How to Pick Survivors

The Gemini CLI shutdown was no one-off: iFlow, Roo Code, Cascade, and Phind died in 2026 too. Get the dated casualty list, survivor traits, and exit checklist.

The Local AI Workstation, 2026 Edition: DGX Spark, Ryzen AI Max, or a Mac?

DGX Spark, Ryzen AI Max 395, Mac Studio, or used 3090s? The 2026 local LLM hardware guide: bandwidth vs capacity, three priced builds, and when local wins.

Industry Analysis14 min read

Goodbye, Llama? Muse Spark and Meta's Proprietary Pivot

Muse Spark is Meta's first proprietary model since Llama. What it is, what happens to Llama open source, and how teams standardized on it should hedge.

Industry Analysis18 min read

OpenAI's Agentic Stack: From Research Papers to Dev Day to What You Can Build

Every OpenAI Dev Day decoded, 2023–2026: the agent primitives that survived, what the Assistants API sunset teaches, and what to build on without whiplash.

Fleet and Swarm Agentic Workflow Architectures in 2026

Compare fleet and swarm agentic workflow architectures in 2026: durable state, bounded delegation, isolated worktrees, evaluations, permissions, and costs.

The Mid-2026 Frontier Scorecard: Fable 5 vs GPT-5.6 vs Gemini 3.1

Which is the best AI model in 2026? Claude Fable 5, GPT-5.6 (Sol), and Gemini 3.1 scored for real agent work — plus the open models closing the gap fast.

The Test Harness, Reinvented: QA for Agentic Software

The test harness in software testing, defined in 49 words — then rebuilt for AI agents: sandboxes, replayed tools, trajectory checks, and budget caps.

Tools & Reviews19 min read

The Best Agentic AI Tools in 2026, Tested by Daily-Driver Standards

Agentic AI tools ranked by daily-driver testing: coding CLIs, IDEs, workflow platforms, frameworks, and the operating layer — with a rubric you can rerun.

Agentic Foundations21 min read

How AI Agents Actually Work: The Loop Behind the Magic

Artificial intelligence agents are a loop: a model deciding, tools acting, results feeding back. See a real annotated trace, then build one in 50 lines.

The Open-Source AI Agent Stack Serious Builders Run in 2026

What open source artificial intelligence really means, and the agent stack that runs on it in 2026: models, runtimes, orchestration, MCP, and three recipes.

Agentic Foundations19 min read

What Is an Agent Harness? The 2026 Field Map of AI Coding Harnesses

An AI agent harness turns a model into a working agent. Get the practitioner's definition plus the 2026 field map: survivors, casualties, and the new wave.

The DeepSeek Effect: Open Models and the Collapsing Cost of Running Agents

DeepSeek collapsed the cost of running AI agents. We price one real workflow across three tiers — down to $0.14/M — and show you which steps to re-route.

The Subscription Squeeze: Copilot Goes Metered and the Flat-Rate Era Wobbles

The GitHub Copilot pricing change swapped premium requests for metered AI credits on June 1, 2026. Why flat-rate AI plans are wobbling — and how to defend.

Agentic Foundations15 min read

What Is Agentic Coding? From Autocomplete to Autonomous Engineering

What is agentic coding? Get the practitioner's definition, the August 2026 tool map, core team practices, and the honest anti-patterns that burn teams.

Open-Source Coding Agents: Aider, Cline, OpenCode, Goose, and Friends

Field review of six open source coding agents — Aider, Cline, OpenCode, Goose, OpenHands, Crush — with maintenance health, BYOK math, and honest picks.

Agentic Browsers in 2026: Atlas, Comet, and the Injection Problem That Won't Die

Atlas, Comet, and Dia can browse and act for you. What agentic browsers do well in 2026, why prompt injection may never be solved, and a safe-use playbook.

Context Engineering: The 2026 Playbook for Agents That Don't Forget

Context engineering keeps agents sharp past turn 30. See what actually fills the window, six techniques with real configs, and a one-week adoption plan.

Harness Engineering: The Discipline Your Agents Were Missing

Harness engineering is why one team ships clean agent PRs while another babysits loops. Learn the six subsystems, day-one practices, and a maturity ladder.

Ox Alpha: The Free Stealth Model Nobody Will Claim

Ox Alpha appeared free and anonymous on August 20, 2026 — 1M context, no maker named. The benchmark that collapsed, the GLM fingerprints, what's safe to send.

Agentic Foundations16 min read

AI Agent Frameworks in 2026: How to Actually Choose

Skip the listicles. A working decision guide to AI agent frameworks in 2026: a taxonomy, an eight-check rubric, a decision tree, and when to use none at all.

Tools & Reviews13 min read

Muse Code: Meta's Terminal Coding Agent, Reviewed

Muse Code reviewed: Meta's beta terminal coding agent, the Muse Spark 1.2 engine, its 59.3% DeepSWE standing, open questions, and how to trial it safely.

Claude Code vs Cursor in 2026: The Agentic Coding Shootout

Cursor AI code editor or Claude Code? We compare autonomy, review ergonomics, and real heavy-user pricing math, then give verdicts by persona. Updated for 2026.

Gemini CLI Is Gone: The Antigravity Era of Google's Agentic Coding

Google shut Gemini CLI down on June 18, 2026, and CI pipelines broke overnight. What happened, how Antigravity CLI replaces it, and the 15-minute migration.

Hardening MCP in Production: The Government-Guidance Era

The June 2026 government CSI made MCP security official. Get the threat classes, a hardening checklist mapped to the guidance, and the new auth upgrades.

The Agent Stack17 min read

Open-Weight Models That Can Actually Drive a Harness

Which open-weight models can actually drive a coding agent in 2026? We define agent-fitness, profile DeepSeek V4 to Kimi K3, and map serving and hardware.

Kimi K3, GLM-5.2, DeepSeek V4: The Open Models Crowding the Frontier

Kimi K3, GLM-5.2, DeepSeek V4, Qwen3-Coder-Next: verified figures, prices, deployment lanes, and how to pick the best open source model 2026 for agent work.

Tools & Reviews13 min read

UiPath and the Agentic Pivot: What RPA's Reinvention Means for Automation

What is UiPath in 2026? The RPA leader's agentic pivot explained: Agent Builder, Maestro, an honest RPA-vs-agents comparison, and who should buy — or skip.

Talk to Your Agents: Voice-Driven Development Grows Up

Voice-driven development grew up: push-to-talk hotkeys, local GPU speech-to-text, and voice-to-spec pipelines. Where dictation beats typing, plus a setup guide.

Product16 min read

Ten Assistants, One Boss: Running Multiple AI Coding Agents Without the Chaos

Running Claude Code, Codex, and Kimi side by side? Manage multiple AI agents with one searchable archive, fleet health alerts, and local token metering.

Agentic Foundations16 min read

What Is an Agentic Workflow? Anatomy, Patterns, and Real Examples

Learn what an agentic workflow is: the seven-stage anatomy, six core patterns, 11 real examples, and when to skip agents — from a team that runs them daily.

Agentic CI/CD: Self-Healing Pipelines and Git When Machines Commit

Agentic CI/CD runs both ways: agents heal failing pipelines, and pipeline gates govern machine commits. Get the playbook, git rules, and adoption plan.

Agentic Foundations24 min read

Agentic Software: How AI Agents Are Turning Code Into Colleagues

What a software agent is, how agentic software actually works, and how to adopt it without chaos — architecture, lifecycle, SDLC patterns, and governance.

The Agent Stack18 min read

MCP Explained: The Model Context Protocol for Power Users

Anthropic MCP explained for power users: how the Model Context Protocol works after the 2026 stateless spec, real client configs, security, and server builds.

The RL Environment Gold Rush: Why Agent Gyms Are the New Training Data

RL environments are the new training data. Why labs pay for agent gyms, who sells them, and how reward hacking and benchmark contamination could sour the rush.

Subagent Orchestration: Fleet Patterns for Daily Drivers

Subagent orchestration without framework theory: five fleet patterns — worktrees, planner/worker, skeptic pairs, swarms, background agents — with real setups.

Spec-Driven Development: From Vibe Coding to Contracts Your Agents Can Ship

Vibe coding broke at review time. Spec-driven development fixes it: a four-artifact stack, one full worked example, real tooling, and metrics that prove it.

Industry Analysis13 min read

When SWE-bench Saturates: Reading Agent Benchmarks in 2026

SWE-bench Verified is saturating — open models post 78–93%. What scores still predict, how vendors dress them up, and a checklist for reading agent benchmarks.

Industry Analysis19 min read

Understanding Anthropic: The Company Betting Everything on Agents

Anthropic explained for builders: the founders, the safety strategy, Claude Fable 5 and Mythos 5, MCP, real critiques, and how to bet on the agent-first lab.

DeepSeek V4 Migration Guide: Life After V3 and R1

DeepSeek deprecated V3 and R1 on July 24, 2026. Migrate to DeepSeek V4 Pro or Flash with real config swaps, an eval-first sequence, and a rollback plan.

Industry Analysis18 min read

Token Plans Decoded: What Heavy AI Users Actually Pay in 2026

AI subscription plans decoded for heavy users: Claude Max, ChatGPT Pro, Copilot's new AI credits, Chinese flat plans, and the math that picks your stack.

Tools & Reviews17 min read

The Anthropic API and Console: A Builder's Guide

Master the Anthropic Console: mint an API key, make streaming Claude calls in Python and TypeScript, cut costs with caching, and ship a small agent service.

Chinese AI Models for Agent Work: DeepSeek, Qwen, Kimi, GLM

Kimi K3, GLM-5.2, DeepSeek V4, and Qwen3-Coder-Next: a lab-by-lab guide to Chinese AI models for agent work, covering capability, licenses, access, and trust.

GLM-5.3: Zhipu's Post-Training Leap and the Two-Week Weights Wait

Zhipu shipped GLM-5.3 on August 14 through the GLM Coding Plan, with weights two weeks out. What the vendor claims, what's verified, and how to try it today.

The Chinese CLI Wave: Qwen Code, Kimi Code CLI, and the $3 Coding Plan

Qwen Code, Kimi Code CLI, and the Z.ai GLM Coding Plan reviewed for August 2026: real prices, quotas, Claude Code wiring recipes, and a calm trust checklist.

Tools & Reviews20 min read

Claude Code: The Power User's Field Guide

Master Claude Code beyond the basics: CLAUDE.md discipline, hooks, subagents, headless CI runs, and cost control — the field guide daily drivers bookmark.

Claude Fable 5 and Mythos 5: Inside Anthropic's Mythos-Class Era

Anthropic split the frontier in two on June 9, 2026: public Claude Fable 5, gated Mythos 5. What Mythos-class means for agent builders, minus the hype.

Train Your Own Mini-Model: From nanochat to a Fine-Tuned Daily Driver

Train your own LLM in 2026: QLoRA fine-tunes with Unsloth, distillation from open teachers, or a $100 nanochat run. Real configs, costs, and honest limits.