Researched and written with 312,891 tokens (in 10 · out 32,222 · cache-read 66,228 · cache-create 214,431) across 5 Claude calls

Coding agents

• v2.1.307 and v2.1.308: Claude Code adds /claude-api api-version to migrate API code to version 2026-10-21 and a claude_code.safety_event OpenTelemetry event for every tool call that a permission rule, auto mode, safety check, hook or sandbox stopped, prompted or sent back, which makes agent guardrails observable.
• 0.167.0: Codex opens an agent command center with codex agents --no-daemon, and both it and the session picker now mark unread results.
• Claude Code sends your email address to websites: a bug report says Claude Code puts the user's email in the User-Agent header during web searches, and commenters tie it to spam and phishing and say Codex doesn't do this.
• Show HN: Stroq: a local audit (npx @stroq/cli) that reads your last 30 days of Claude Code and Codex sessions and reports how many commands came from text the agent had read (a prompt-injection signal) and which keys left your machine; commenters say eBPF would cover more adversarial cases and dev work should run in a VM anyway.
• Prompt of the day: asking the agent "How do we make this change fully verifiable?" makes it verify its work by signing up for real services and clicking through the frontend.
• Gave Claude Code the Chrome DevTools MCP and a broken page: a report that Claude Code found two bugs on a broken page through the DevTools MCP without the user opening DevTools.
• Built with Claude Code: Still lets Claude Code drive browser tasks in background tabs: a Claude Code-built tool that lets the agent run browser tasks in background tabs, so it doesn't take over your active window.

LLM context and local inference

• "You Don't Need Long Context": NVIDIA reports that agents can finish long-horizon tasks without carrying the full conversation, using frequent compaction that summarizes progress and retrieves details on demand, even resetting context after every action, which cuts inference cost.
• Qwen3.8-27B Benchmark: compares 6 inference engines, GGUF vs NInfer, at 256K context on both speed and quality, which is useful if you pick a local serving stack for long-context agent work.

Agent tooling and skills

• vercel-labs/agent-browser: a native Rust CLI where snapshot returns the page as a compact accessibility tree with refs like @e1, and the agent acts on a ref and re-reads the page to verify, one shell command at a time.
• Leonxlnx/taste-skill: an Agent Skill (v2.5.0) that pushes coding agents toward art direction, reference images and a browser check instead of generic AI-looking sites.
• google/skills: Google's Agent Skills for its products, including Google Cloud, installed with npx skills add google/skills into your chosen agents and scope.

Agentic SaaS

• Launch HN: Rivet (YC W23): long-lived Actors with in-memory state, SQLite, WebSockets and scheduling, one per agent, user or session, claiming 4.99 ms p99 cold starts and 72 KB RAM each; commenters find the UX confusing.
• Cosyra: launch and steer agents on your own machines from your phone.
• Vibeconferencing: puts your Claude Code session into a Meet, Zoom or Teams call.
• Glasshand: gives an agent a real iPhone and Mac to operate.
• Okapi: AI agents that fix API breaking changes in your repos.