Field notes from
production agents.

The Monday-morning bill-shock story: one team's agent burned $1,900 over a weekend, and the per-agent budget setup that stopped it.

From 'the agent can call everything' to a tight allowlist: what broke, what stayed, and what a real support agent's allowlist ended up containing.

Three incident patterns — a retry loop, an exfiltration attempt, a runaway model bill — where a kill switch was the only thing fast enough.

The approver-fatigue problem: how one team tuned their queue with risk tiers, auto-approve lists, and batching so approvals take seconds, not hours.

One shared API key means one leak rebuilds your whole fleet's access. A walkthrough of per-agent credentials: issuing, scoping, and rotating them without the 2 AM scramble.

Denylists fail because you can't enumerate everything an agent shouldn't do. How a fail-closed proxy flips the model: every action evaluated, default deny, and what the agent sees when it hits the wall.