Things I built, and what they had to survive.

Multi-agent engineering method

Airtable · 2026

• 384Pull requests, every one reviewed
• 251Sub-agent dispatches in the final month
• 1Accountable human at every merge and grant

Running AI agents like an engineering team

The warehouse was not built by working longer hours. It was built with an orchestration loop: one standing session that never writes code, worker agents building in isolated git worktrees, and a merge gate where every pull request gets a fresh-context review, a durable verdict, and a human acknowledgement before it lands.

384 reviewed pull requests in eleven weeks, none merged without a human gate. The method is documented internally and reused on other builds.

Performance check-in redesign

Airtable · 2026

• Company-wideEvery employee covered by the new check-in
• Drafts onlyThe agent never publishes, scores, or ranks
• ZeroPremature notifications during rollout

Rebuilding a quarterly check-in as a reflection agent

I led the 2026 rebuild of Airtable's quarterly performance check-in, replacing a form and approval workflow with an agent that helps each person assemble their own account of the quarter. I set the trust framing, wrote the specs, built the agent skill and its home and Slack surfaces, and sequenced the launch so nobody heard from the system before they were told what it was. In parallel, I kept the legacy check-in running through its live cycle.

The check-in is now an agent employees run themselves, a weekly digest through the quarter and a guided reflection at the end, in a home surface and in Slack, and every output is a draft the person owns.