Researched and written with 264,861 tokens (in 8 · out 36,754 · cache-read 51,896 · cache-create 176,203) across 4 Claude calls

Claude Code and LLM mechanics

• Anthropic shipped message threads (beta): Claude Code now sends only the new messages each turn, cutting time to first token by about 35% on 500k to 1M token requests, which matters for long-context agent sessions.
• Claude Code v2.1.318: adds a record of your answer when Claude Code asks to enable session-written mods, assume_role on mantle upstreams of the Claude apps gateway (cross-account AWS roles, per-developer cost attribution), and an effort option on a mod's $.model.fork.
• AGENTS.md ignored when CLAUDE.md exists: Claude Code loads only CLAUDE.md by default when a repo has both, so cross-tool instruction files can silently go unread.
• Game-project pushback rule loosened in 2.1.317: the Opus 5.5 prompt rule meant to stop recreating specific existing games also fired when you merely named one to describe mechanics, and 2.1.317 narrows it.
• Haiku 5.5 ignores thinking "off": it kept reasoning with thinking set off, and with it fully disabled wrote the answer first and rationalised afterwards, so don't trust the toggle to mean no reasoning tokens.

Other coding agents and IDEs

• Codex rust-v0.167.0-alpha.5: new alpha tag of the Rust Codex CLI; the feed captured no changelog content, so check the release page for details.
• GitHub Copilot in JetBrains Air: delegate and review coding tasks from Air through the Agent Client Protocol (ACP), another sign ACP is becoming the common editor-to-agent interface.
• Continue a Copilot app session in VS Code: moves an agent session between the Copilot app and VS Code while keeping conversation, code changes and working context.
• Dependabot supports Agent Package Manager: Dependabot now opens PRs to update APM-managed agent skills, prompts and instructions, treating them as versioned dependencies.
• Team of AI agents in VS Code supervised by Claude Code: a French-language field report on running several agents in VS Code with Claude Code as supervisor.

Agent harness practice

• Paper on Matt Pocock's /retro skill: researchers wrote a paper assessing how effective his /retro skill is, which is rare outside evidence for a skill-based workflow.
• Harness self-evolution needs a way to check changes: the paper's point is that a harness change needs a check on later tasks to show it helped, with revision or removal if not, so /retro changes themselves need a retro.
• Per-agent wiki to switch harnesses: each agent builds a wiki over time so conversation history can be dropped and the harness swapped, with skills like /implement-spec and /retro on call.
• Orchestrator thread plus software factory: a long-running local orchestrator per domain that drives a factory of sandboxed implementers over the CLI, manages threads, runs retros and builds cron jobs, with the human speaking plain language.
• How many attempts did your agent take?: a correct answer can hide cost, since retries, slow tool dependencies, extra context and larger models each add calls, tokens or spend, so track attempts per answer.
• kwant: AI bots that check each other's work before reporting "done", a verifier-agent pattern.

Agentic SaaS

• Claude auto-connects company plugins: Team and Enterprise admins can pre-configure plugins so teammates connect automatically on first use, with no per-plugin sign-in.
• Plugin auto-connect for MCP builders: MCP servers that accept Claude as a trusted issuer get auto-connect at login, so plugin authors should add that support to cut onboarding friction and re-authorisation.
• Peji: an AI chat paired with a real browser you control.
• Scrawl: lets you show your AI agent a problem with a visual frame.
• Stageshow: a free skill for launch videos that turns UI screenshots into a story, reportedly 22,627 bookmarks in 37 hours.