AI agent & Infrastructure engineer. building loops, graphs and harnesses for autonomous production and scaling using AI I turn research questions into working software: measurable, inspectable, and useful past the demo.
- Building Metaorcha: one goal in, a verified multi-agent run out. Metabot on top, in progress.
- Production DevOps and AI engineering on live, high-traffic CRMs. The method
- Writing The Harness Layer, field notes on agent harnesses.
More: commit-city · RP2A · sdca-raqqr
- 5/5 multi-protocol runs verified live, best wall clock 13 s · Metaorcha
- 41% fewer agent steps than a fixed budget · field note 009
- 25 typed MCP tools behind a production voice-ordering agent · RP2A
Stack · Python, TypeScript, SQL · MCP, A2A, Claude Code hooks · FastAPI, Next.js, Postgres, Supabase, Vercel · OpenTelemetry, Bayesian optimization
Upstream fixes
| date | where | what | outcome |
|---|---|---|---|
| Oct 31 | agent-check | v1.1.0: now flags comment-triggered checkouts, and documents a fix for every finding | released |
| Oct 31 | guardrails-kit | fail-closed Claude Code hooks for team repos, with tests | released |
| Oct 19 | tmux #5775 | ^ matched again mid-string in format substitutions, since 2019 |
applied upstream 10 h later, regress test credited |
| Oct 18 | blackbox_exporter #1715 | target URLs percent-decoded twice since 0.29.0 | open, awaiting review |
| Oct 16 | huggingface_hub #5180 | hf skills add --force replaced a skill with a symlink to itself |
fixed in-house in v2.3.0, report credited |
| Oct 15 | ECC #3591 | SessionStart hook crashed in every project on plugin installs | merged Oct 29; on Windows 0/27 → 26/27 hooks passing |
| Oct 15 | vercel-plugin #343 | session-start context overran Claude Code's hook output cap | open; a Vercel engineer extended it on the branch |
| Oct 15 | BMAD-METHOD #3094 | update drops the config of a module kept without a source | open issue |
On carving your own path
Side quests used to be what you did after work. I made them the work, and made it respectable.
The method was simple: doubt every source except my own judgment, then keep retraining how I decide things, scoring each move by how much it widened my exposure to lucky breaks. Run that loop long enough and your choices converge on the handful that actually pay off, while a small share of every attempt still goes to wandering on purpose.
Where it led: companies with nobody on payroll, built one day at a time as my way out of a class that was never meant to move.
If you want a path that's genuinely yours, my one piece of advice is to defend your mind like it's production. Trends flip. Focus erodes. Someone will always be trying to slip their narrative past your filters and edit your worldview for you. Don't accept the patch.
Outlast the noise and you win.
metaorcha.ai · X · It's all relative.