Ten Things to Wire Up Before an Agent Touches Production
A developer deployed a customer-support agent that got stuck in a retry loop with a CRM tool. No hard circuit breaker. It spent six hours overnight repeating the same broken call while he slept, and he woke up to roughly a $4,200 OpenAI bill for doing nothing useful. That's the failure mode. Not a dramatic hack — a boring loop with no external stop. Roughly 95% of enterprise generative-AI pilots…
1. A developer deployed a customer-support agent that got stuck in a retry loop with a CRM tool, costing around $4,200 due to OpenAI tool calls.
2. 95% of enterprise generative-AI pilots in 2025 delivered no measurable return, showing teams often optimized the brain but skipped the nervous system.
3. The shift from reading to doing in agents, where they perform actions like writing to a database or calling APIs, is where pilots commonly fail.
4. LLMs should be treated as fallible kernels, not magic boxes, requiring checks, limits, and rollback mechanisms.
5. Agents break the assumptions of determinism, leading to different actions on different days for the same input, creating a "ghost debugging" issue.
6. Design for a tool-call failure rate between 3% to 15% as a normal state rather than an exception, as loop cost grows with length, not request count.
7. An agent looping produces token usage that grows quadratically, making 20-step loops far more expensive than 10-step loops.
8. Each tool connected to an agent widens its blast radius, weakening its safety controls.
9. Four controls must be wired before touching agent logic: circuit breaker, hard cap on retries and loop iterations, deny-by-default allowlist, and confirmation gate for destructive actions.
10. External controls, not the prompt, protect against catastrophes; they don't make a weak agent good but provide safety to improve it in production.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
