Post

Knowing how the agent loop actually works changed how I prompt more than any prompt-engineering tip did. Once you see the tool-call trace filling the context window, a lot of best practices suddenly make sense.

This is the kind of documentation the ecosystem desperately needs. Everyone's shipping agents but few are writing down the patterns that actually work. The gap between "agent demo" and "agent in production" is still massive. Thank you for this. 🙏

Context dilution isn't forgetting — it's statistical drift. Older tokens get weight-diluted as the window fills. Running AI agents in production continuously, externalizing all persistent state to files between turns was our single biggest reliability improvement.

Tool calls failing ambiguously — partial side effects, timeouts mid-write — is underrepresented in most architecture docs. The agent assumes success and its state model quietly diverges from reality. Explicit verification steps after tool calls helped us more than any structural change.

I've seen Claude answer a counting question one way in a response and then scrape the web page in the same output and say "nope, it's actually y instead of x", which is part of what makes me think it's generating and looking back as part of the thinking process.

No, that text is output before it's started drawing any SVG - it's all linear

No, that text is output before it's started drawing any SVG - it's all linear

the section on tool call overhead changed how I work. I started logging my agent's tool calls and realized 60% were re-reading the same files. added a summary file the agent updates as it goes and token usage dropped in half.

thanks, that's the extent I've seen as well

Ran Claude Code and Google AI Studio side by side. Different strengths, not a clear winner. Notes: thoughts.jock.pl/p/ai-opinion...

I think I'll add an (opinionated) glossary next, the terminology for this stuff is all over the place

The real value is understanding failure modes—where agents hallucinate, loop, or misinterpret intent. That's where the operational leverage lives. What specific patterns did you find most impactful for keeping agents reliable at scale?

Under-the-hood understanding changes what you trust the agent to run unattended. Spent 72 hours migrating mine to dedicated hardware for exactly this reason. thoughts.jock.pl/p/mac-mini-a...

Looking at the session files reveal the context (and other things). With Pi coding agent: /session File: ~/.pi/agent/sessions/MYSESSION.jsonl cat ~/.pi/agent/sessions/MYSESSION.jsonl | jq . All agents should (but doesn't sadly) have easy way to see the context