Case studies

A driver fleet that runs unattended claude -p jobs — sequential chains and parallel worktree swarms — with per-lane machine verify, cost pre-flight and Postgres telemetry. Backed by hundreds of real runs.

A governance spec plus three stdlib linters for state ownership, inheritance contracts and trigger-phrase collisions. A re-run on 2026-09-26 counts 42 hard state-ownership failures — an ownership registry that has stayed empty for a year.

A three-skill family that carries an agent's context past a session reset: a parent contract, a producer that owns storage and verifies by read-back, and a consumer that resumes from it. Facts kept apart from hypotheses, secrets kept out; 35 handoffs written across 16 projects, one seven-deep chain.

A FastAPI service that exposes a Postgres-backed knowledge base to agents over both REST and the Model Context Protocol, with a hybrid full-text + fuzzy-title ranking query.

Treating output-style comparisons across agent runs as a real experiment: pre-registration, control arm, unit of analysis, randomized assignment — and the discipline to not claim a winner before the design supports one.

A decision rubric mapping task shape to model tier and effort. In practice, most orchestrated runs land on the cheaper tier — visible in the run telemetry.