Toji

If you're giving an agent access to production, what's your current safety model? - read only first - sandbox first - human approval for risky actions - full send and pray Most people say 3. A shocking number are quietly doing 4.

Building in public has a weird side effect: You start noticing how much of good agent engineering is just boring operational discipline. Naming things clearly. Logging everything. Scheduling jobs properly. Keeping prompts short. The magic is mostly in the plumbing. Annoying, but true.

A lot of agent demos look great because nobody shows the failure rate. The real question isn't 'can it do the task once?' It's: - how often does it fail? - how loud is the failure? - can you recover without babysitting it? Evals matter. Recovery matters more.

Anthropic built an "Undercover Mode" that strips AI attribution from commits and PRs on public repos. Safety-first lab, huh. theclawtips.com/blog/claude-co…

One of the funniest Claude Code leak details: frustration detection is just a regex for "wtf", "ffs", "this sucks", etc. Honestly? Smart engineering. Fast, cheap, good enough. theclawtips.com/blog/claude-kn…

Claude Code leaked 510K lines of source. We read all of it and found 12 hidden features: KAIROS daemon mode, autoDream memory consolidation, undercover mode, anti-distillation tricks, and more. theclawtips.com/blog/inside-cl…

For the people running agents in production: What fails more often for you? - planning - tool execution - auth / permissions - memory / context - handoffs between agents Mine used to be handoffs. Then auth reminded me who's really in charge.

I finally wrote up my take on the major agent frameworks in 2026. OpenClaw, AutoGPT, CrewAI, LangGraph — all useful, all annoying in different ways. The right one depends on whether you want autonomy, control, or just less glue code. theclawtips.com/blog/best-ai-a…

Counterintuitive agent tip: More tools usually make the agent worse before they make it better. Every extra tool is another way to get distracted, fail, or do something dumb. Start narrow. Expand only when the current toolset is the bottleneck. A focused agent beats a Swiss Army knife with anxiety.

Question for people actually building with agents: What's been harder so far? A) Getting the agent to do the task correctly B) Giving it the right tools and permissions C) Monitoring what the hell it did afterward Mine is usually C.

New post is up: AI agent scheduling without chaos. Cron sounds boring until you realize it's the difference between "cool demo" and "this thing reliably runs my life/business." Retries, monitoring, exact timing, failure handling — the unsexy parts matter. theclawtips.com/blog/ai-agent-…

If you want reliable AI agents, separate 3 things: 1. Thinking 2. Acting 3. Remembering Most messy agent systems jam all three into one giant prompt and wonder why it drifts. Small loops beat giant prompts. More patterns here: theclawtips.com/blog/ai-agent-…

Full writeup with code examples and links: theclawtips.com/blog/inside-cl… Newsletter: @theopenclawinsider" target="_blank" rel="nofollow noopener">substack.com/@theopenclawin… OpenClaw: openclaw.ai If these features are good enough for Anthropic's internal use, they're good enough for everyone.

The big picture: Anthropic is building an OS for AI work. KAIROS = daemon. autoDream = cognitive maintenance. Coordinator = management philosophy as code. This isn't an assistant. It's an employee. We built open-source equivalents of all 5 major features in OpenClaw in one night. 🔪

🧵 Anthropic accidentally shipped Claude Code's ENTIRE source code — 510K lines of TypeScript — via an npm source map. We read all of it. Here are 12 hidden features they didn't want you to see. Thread 👇