agents

I had Claude write nearly all of a two-layer eval harness for a client's RAG app. The design calls are mine, and they're what make it trustworthy: the judge model's own statement is stored next to the verdict my code derives, so a bad grade contradicts itself somewhere I can see. What I measured, what I rejected on the numbers, and where the numbers still can't help me.

I built Karpathy's LLM wiki pattern into Command, my Claude Code operator: 134 provenance-tagged pages, a grep gate that blocks fabricated quotes, a nightly digest pipeline that was reset after a cheap model wrote a wrong verdict into it, and a human-gated publish wall.

The key design decisions in Command, my autonomous Claude Code operator: why the watcher is deterministic and free, why it authors per-repo config instead of loading it, why every fix carries a pre-registered success criterion, and what I want to see from it next.

Prose design specs drift the moment nobody's looking, and coding agents now read them as ground truth. The two-stage CI guard in chat-arch treats every palette-table row as a falsifiable claim: what it catches, the contrast bug that earned its own token, and the geometry drift it sailed right past.

71 invocations of the adversarial-subagent-team pattern in Claude Code and Cowork, verified against the cited PRs, commits, and arXiv references. Across six headline claims: three confirmed clean, one confirmed with caveats, one rejected on falsification, one not supported by the data.

Anthropic's latest harness research reveals two underappreciated patterns: why agents can't evaluate their own work (and what to do about it), and why you should be actively removing complexity from your agent harnesses as models improve.

Subagents, skills, CLAUDE.md orchestration, headless execution, and state management patterns for Claude Code multi-agent systems. 50 claims verified against official Anthropic docs.

How I built ShopForge, a production AI agent with three-tier memory, Bayesian confidence scoring, and skill evolution, and what 10 days of real Etsy operations taught me about self-improving systems.