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.
Court filings reveal Character.AI's Chapter 11 was the second step of a Disney deal to transfer 60+ staff and a technology license while the remaining shell retains all legal liabilities over chatbot safety.
Browser automation tool is GA in Microsoft Foundry: managed Playwright browsers, MCP toolboxes, private-network reach and human takeover for sites with no API.
Beijing is weighing export licensing for AI models and chips just as the shared open-weight commons nears an estimated $1B in donated training compute.
A single-commit GitHub repo cloned the popular blender-mcp project, wrapped it in a polished 'MCP Pro' README, and — per a teardown — dropped StealC on enable.
Qwen Code v0.25.3 goes stable with snapshot-isolated read-only child agents, CSI-backed Kubernetes workspaces that persist across sessions, and in-session approval-mode switching that keeps history and drafts.
Vercel AI Gateway now groups requests into sessions with cache hit rate, per-session cost, and coding agent tracking. Session affinity keeps turns on one provider by default.
The FT reports Anthropic is consulting law firms on negotiation tactics for Claude agents that bargain on humans' behalf. The sourcing is thin, but the operator questions are concrete: spending caps, signed mandates and liability when an agent clicks agree.
Yelp, Delta, Chase and Schwab are turning AI agents away. The operator playbook: sign your identity, honor robots.txt, back off on 429s, and sandbox everything.
OpenCode confirmed a breach tied to a large-scale distillation attack exposed stored OpenAI, Anthropic and Google keys for 1.4% of Console users with linked providers. Here is the operator response: rotate, cap spend, audit billing.
DeepSeek opened Search and Fetch APIs at $0.75 per 1,000 calls each, free in Claude Code and Codex until October 31, powered by its in-house search engine.
Meta stripped the Commercial Terms clause experts read as a license to scrape news sites for AI training, narrowing it to ads and business products days before it was scheduled to take effect.
DeepSeek Harness v0.2.1-rc.1 adds Claude Code and Codex bundles, agent-built plugins, a Git Worktrees plugin, and shared AGENTS.md for multi-harness fleets.
NVIDIA RTX Spark laptops shipped October 16 from six vendors starting at $2,599, pairing a Grace CPU with a Blackwell GPU and up to 128GB of unified memory. Here is what developers can run on one machine, what Junie Local delivers at 55 tokens per second, and where a week of Verge testing tempers the pitch.
NVIDIA's open 120B Nemotron 3.5 Super VL unifies text and vision for specialized agents, with NeMo tuning, NIM serving and SageMaker day-zero deploy options.
Claude Code v2.1.302 skips credit-billed model suggestions, adds an idle compaction toggle to /config, and enables fork subagents by default in print mode, SDK, desktop, and VS Code.
Anthropic says Claude Mythos 5 designed quantum circuits beating published records on 7 of 8 benchmark tasks, cutting gates up to 17x via a new divsteps algorithm.
Three October 16 paths to Reflection Beam: Google Cloud cache-aware serving, SageMaker JumpStart and HyperPod, and Ollama Cloud for a limited time with stated zero data retention.
Cloudflare announced a continuous evaluation system, already in internal use, that measures Verified bots on robots.txt compliance and HTTP 429 response, producing explainable trust bands instead of permanent scores. Public Radar reporting and customer-selected trust thresholds are planned; no release date was stated.
Mistral AI has released Timestral, an open-weight time-series foundation model family from 5M to 209M parameters. Engineered for zero-shot forecasting on commodity CPUs, Timestral reports leading benchmark scores across GIFT-Eval, TIME, and fev-bench under a Modified MIT license.