Researched and written with 320,787 tokens (in 10 · out 38,382 · cache-read 63,690 · cache-create 218,705) across 5 Claude calls

Coding agents

• v2.1.323: Claude Code adds opt-in experimental Code Mode, where Claude writes JavaScript to call tools; also engines["claude-code"] in plugin.json for plugin version gating.
• 0.168.0: Codex adds model-provider browsing and default selection per new chat, plus features.tui_workspace side conversation view.
• QwenLM/qwen-code: Open-source terminal/editor/desktop/browser agent; 1.0 Preview merges Qoder harness capabilities, includes auto-memory. Credible OSS Claude Code alternative.
• Alishahryar1/free-claude-code: Unofficial proxy routing 12 harnesses (Claude Code, Codex, OpenCode, Aider, etc.) to 61 providers incl. free tiers. Useful for harness/model decoupling; ToS and reliability risk.
• openai/codex-security: OpenAI CLI, TypeScript SDK, and GitHub Action to find, validate, and fix vulnerabilities; scans repos, paths, or git diffs.
• ADEMark v1: Benchmark of agent harnesses on same Mac mini, tasks, and model against bare CLI baseline; most harnesses used fewer tokens than bare CLI, one idled at 16% CPU.
• Prompt change: "owner and long-term maintainer": Recommended prompt now reads "You are the owner and long-term maintainer of this codebase, and its design is up to you." Cheap framing lever for agent autonomy. Same change confirmed here.
• Opus 5.5 builds fun video games: Simon Willison calls October another coding-agent capability inflection point; Boris Cherny RT.

Managed and self-hosted agents

• Claude Managed Agents self-hosted (private beta): Harness, sandboxes, credentials, and session history run in your own cloud. Enterprise data-perimeter unlock. Duplicate thread.
• Claude Meetings: Claude joins team video calls and builds docs, whiteboards, artifacts live; beta on Team and Enterprise.

Multi-repo, skills, and context hygiene

• Home Repo pattern article: Matt Pocock's write-up on one central repo holding skills, steering files, and tooling for agents spanning many repos; includes setup prompt. See intro thread.
• Build once, run on any agent: agent skills: GitHub guide: skills are portable reviewable files, while tools, context, and permissions stay host-specific.
• Manual "dreaming" pass for CLAUDE.md and memory: 10-minute routine to prune and consolidate memory files; counters context rot.
• GoogleChrome/modern-web-guidance-src: Skills embedding web platform best practices and browser compat data to steer agents off legacy patterns.

Agent orchestration UIs

• Conductor is back: Relaunch after sudden shutdown; HN commenters who moved to T3 or the OSS fork see no reason to return, distrust Meta with coordination data.
• T3 Code vs Claude Code/Codex: Pitch: 6 Claude, 1 Codex, 1 Cursor account in one open-source app. Chris Titus also reports it beats Codex and Claude Code now.
• 8 agents in parallel, tracking is the bottleneck: Building is fine; supervising parallel agents is the pain. Same gap Conductor and T3 target.
• 5 Claude Code subagents for job applications: Field report on subagent fan-out, what worked and what broke.

CLI/TUI

• simonw/llm: CLI and Python library for prompting OpenAI, Anthropic, Gemini, Qwen, DeepSeek, local models and more from the terminal.

Vibe-coding

• Tool showing how a vibe-coded app actually works: Comprehension tooling for agent-generated codebases.