The Agent OS: a harness to survive a $100M to $400M year

Supabase took six years to reach $100M ARR, then jumped from $100M to $250M in six months. Data engineering stayed at four people. This is the field report on how: not by hiring, but by climbing a ladder from prompt engineering to context engineering to harness engineering.

We bet early on a machine-readable context layer: versioned docs describing the warehouse, 40 metric definitions, 67 pipelines and 210 feature-coverage entries with YAML frontmatter, curated by a three-person data intelligence team and kept current by agents that audit Slack, Notion and git weekly. On top of it we built the harness: the deterministic layer of skills, permission gates, channel overlays, triage logs and evals that turns a model into a teammate. Today an AI teammate triages our Slack channel, monitors pipelines, writes the weekly executive business review and opens draft PRs, governed by a constitution and an agent registry.

I will show what broke along the way (an agent hallucinating customer feedback tickets, cost blowups, our own context docs drifting) and the guardrail each failure produced. You leave with a harness blueprint and one argument: your context layer, not your dashboards, is the asset AI will multiply.