Working in an AI-first company in 2026 (without the hype and the fluff)
Working in an AI-first company in 2026 (without the hype and the fluff)
At a title-company software team, Enrico Teotti describes an internal product record that gives coding agents access to interview excerpts, decisions and prototypes. For a partial-receipts feature, customer examples led to four interface options; a working prototype exposed a way to double-count a deposit before implementation. A September 11 plan joined the expected behavior to five implementation units. Four specialist review agents then examined the code; adversarial review caught a retry that could post the same payment twice after the interface closed. The team changed how the retry identifier persisted and added a test.
A separate ledger-syncing change later worked locally but its background job did not run in production. An earlier review had warned about just this kind of unmonitored post-commit side effect. The warning remained in review output rather than entering the documented solutions that the next agent searched before planning. Teotti also found that he could learn the codebase faster than the title industry, and that running two feature builds in parallel exceeded his capacity for product decisions.
The company and some details are intentionally obscured, so the feature, incident and customer evidence cannot be independently joined to public PRs or a release. The author reports about one hour of agent-active elapsed time for partial receipts, excluding interviews, human review, testing, model bills and later support. The production incident was a different change; there is no same-feature customer retest or controlled productivity comparison.