I build, write, and speak about agentic AI, product architecture, observability, and the future of human-AI workflows.

Right now most of my energy goes into one question: when software starts doing the work, who is watching the software?

The local-first flight recorder and control plane for AI agents. Your agent's calls pass through it; every prompt, token, and dollar is recorded on your machine, and a runaway loop meets a wall that can actually say no.

uv tool install agentic-ledger && agenticledger start

Someone's agent once spent $6,000 overnight without crashing. It just looped. That should never happen to anyone again.

→ agentic-ledger.dev

• Measured, not promised. Performance claims trace to a script you can rerun; cost math is pinned by tests against real provider bills.
• User zero eats the bugs first. Nothing ships until its primary scenario passes under my own hands, and every mess my machine produces becomes a product fix so yours never sees it.
• Small releases, honestly named. Ship what is stable, say what is missing, let the changelog tell the truth.
• Plain words. If a feature needs jargon to sound useful, it is not ready.

Issues and PRs are the best doorbell: the good first issues are genuinely scoped, and outside contributions have landed within hours of asking. If your agent loops in a way my detector misses, I want the transcript.