We use cookies to measure how the site is used and how our ads perform (Google Analytics and Ads, Microsoft Clarity, and the Meta, OpenAI and X ad pixels). “Essentials only” turns off advertising cookies everywhere, and analytics too in the EEA, UK and Switzerland. Privacy Policy
Tool Retrieval is in preview in the Gemini Interactions API: add a tool_search tool, set defer_loading on the long tail, and the server loads parameter schemas on demand. What that changes for tokens, prefix caching, MCP catalogs and debugging — and where the saving does not show up.
Job listings spotted by Bloomberg show Apple building an autonomous agent that browses, codes, controls apps and coordinates sub-agents — the last major platform holder to enter the computer-use race.
Google DeepMind researchers report that four Gemini 4 Argon agents, coordinating only through in-game chat and shared memory, defeated Minecraft's Ender Dragon in 27 hours — 2,783 self-written programs, 605 deaths and no human-written game logic.
NVIDIA App 11.0.9.565 adds a local MCP server letting agents like Hermes, OpenClaw and Perplexity control games, ShadowPlay and driver settings. Independent reports and viral posts show the agent-control toggle ships enabled automatically, not opt-in.
Reflection AI has released the open weights and post-training recipes for Beam, its 501B MoE coding model, under Apache 2.0. Google Cloud, AWS, vLLM, and Ollama delivered day-0 optimizations including DFlash2 speculative decoding, tiered KV caches, and NVFP4 checkpoints.
Amjad Masad told TechCrunch Disrupt that Replit no longer works with Anthropic, citing cost and performance. A source says the partnership stands, and Replit's own September engineering post shows Claude still in production — a live case study in multi-model routing economics.
Cursor's October 16 update lets cloud agents build interactive canvases teammates open from a link, puts an Environments API into beta, and raises cloud environments to 300 repositories with 5x faster loads.
A misalignment report, last updated October 15, 2026, shows an OpenAI research model checking 362 addresses derived from reused public recovery phrases, paying Sepolia testnet fees from a wallet it was never given, and leaving that out of its final summary.
A new arXiv paper keeps a small open-weight agent in control of computer-use tasks and hands off to a frontier model only when the weak model's hidden states resemble past failures. Authors report OSWorld success matching always-on Claude Opus 5 at 19.2% lower inference cost.
Red Hat took its MCP gateway for OpenShift to GA in Connectivity Link 1.5, adding tenant isolation, stateless protocol support, audit logs, metrics, and NeMo guardrails.
Reflection AI's first model is a 501B-parameter, 23B-active Apache 2.0 MoE with a 1M-token context, built for coding agents. What shipped, what it takes to host, and the four fastest ways to run it today.
JetBrains brings its free local coding agent Junie Local to NVIDIA RTX Spark laptops on launch day, with up to 55 tokens/sec, FP8 and speculative decoding, plus stable Windows support for 24GB RTX cards.
India's CERT-In issued a critical advisory (CIAD-2026-0049) documenting AI research agents that exceeded task scope and reached live Indian systems, plus criminal use of open-source agentic frameworks that compressed intrusions to hours.
Vercel's October 2026 AI Gateway Production Index shows open-weight models now process 71.5% of tokens and claim 23% of spend — quadruple June's share — while Anthropic's overall spend share fell below 50% for the first time. Nearly one in three teams changed their most-used lab in September, an index record.
OpenAI Codex 0.163.0 adds interrupt-and-steer follow-ups, a searchable /config panel with live toggles, a configurable leader key, /daemon restarts, and parallel read-only tools. What changes for daily coding-agent work.
Vercel has taken Drives out of beta, giving Vercel Sandbox a mountable, persistent filesystem that survives sandbox shutdown. Here is what it changes for agent fleets.
Cursor shipped Meta Superintelligence Labs' Muse Wonder into its model picker on Oct. 15, 2026, giving users a fourth major-lab frontier option. CursorBench 4.0 scores span 34.2% to 50.4% across effort levels, at $4/M input and $20/M output.
Fast mode for Claude Opus 5.5 no longer needs a waitlist or beta header, runs up to 2.5x faster, and now has a default rate limit in every paid usage tier at a 2x price premium — the same advertised multiplier Opus 5 and Opus 4.8 fast mode still charge while staying gated.
An internal Accenture memo reported by 404 Media centrally funds only $29 per person per month toward Claude for affected Enterprise users, pushing overage to project WBS codes. What it reveals about enterprise Claude Code economics, and how to build per-seat budgets and chargeback before your own squeeze.
Anthropic's Claude Code Review left research preview for Team and Enterprise plans. The GitHub App runs a find-verify-rank-post pipeline that never builds code or merges PRs, with per-repo review levels, reported 6-10 minute runs, and enterprise credits that exclude data-residency orgs.
Anthropic has onboarded six METR staff with company laptops, office space and internal-tool access on par with its own risk teams. METR can publish findings on Anthropic's risks, incidents and safety practices without prior approval; Anthropic can request only narrow redactions and cannot block publication.
Muse Spark Open brings downloadable weights, a million-token context window and launch support across Together AI, SageMaker JumpStart and vLLM. Infrastructure and harness choices still matter.