Running agents in production
Cost, context, evaluation, and the guardrails that keep an agent from doing damage.
- Context and cost, measured instead of guessedLevel 2 · 8 min — Where the tokens actually go, why tool definitions are charged on every request, and the three levers that change the bill.
- How to know your agent got worseLevel 3 · 10 min — A minimal eval harness: twenty cases, a deterministic check where possible, and a rule for when to re-run it.
- Guardrails: permissions, approval, and auditLevel 3 · 9 min — What to hand an agent, what to keep behind a human, and how to reconstruct what it did afterwards.
- The last hour before launchLevel 3 · 8 min — Eight checks that catch the failures which only appear once real people and crawlers arrive.