Wheelhouse and Counting House: Two ways to put models to work
summary
Two production patterns: a 50-agent Wheelhouse software factory that ships about 2,000 commits a week, costs ~$10–12k/month in inference (22 Claude Max accounts, ~$1.5M API-equivalent), and behaves like people with ~18 direct-report agents, doctrine, and blame-free postmortems; and a rule-driven Counting House for high-volume CS that enforces no PII ever reaches a model, forbids model writes, uses agents to emit write-intents executed by Temporal, and rolls agents through off → shadowing → assisting → transacting with eval matrices and cognitive-surrender guards.
AI-written summary · about
chapters (9)
- 0:00Introduction by Gene Kim
- 1:04Two ways to bring in agents
- 3:17Wheelhouse: managing agents like people
- 6:26Fuel costs and living in the future
- 10:35Pacing agents and blame-free postmortems
- 12:14Counting House architecture
- 16:16Choosing models with evals
- 19:43Autonomy tiers and rollout
- 21:13Which house are you in
