Your agent passed every eval. Then the CRM API went down for forty minutes on a Tuesday, and your "autonomous" agent spent those forty minutes confidently inventing customer records it could not reach. The model wasn't the failure. The failure was building an agent that assumes the world answers when it asks.
Every agent is a chain of dependencies wearing a trench coat. It calls your CRM, your billing system, a search API, a document store — and each of those has its own uptime, its own rate limits, its own idea of what an error looks like. A model that hallucinates gets all the attention. An API that returns a 200 with an empty body gets none, and it will quietly corrupt more decisions in a week than a bad prompt will in a month. The honest reliability number for your agent isn't the model's accuracy. It's the product of every dependency's reliability, and that product is always smaller than you think.
So design for the outage, not around it. Every tool call needs a timeout that is shorter than the customer's patience, a retry policy that backs off instead of hammering a struggling service, and — the part teams skip — a defined behaviour for when the answer never comes. What does the agent do when the CRM is unreachable? It should say so, plainly, and stop. An agent that reports "I couldn't reach the CRM, so I haven't touched your record" is a product. An agent that fills the gap with a confident guess is a liability with a friendly tone.
Test it the way you'd test anything else: break the dependencies on purpose. Point the agent at a staging CRM that returns 500s, at a billing API that times out half the time, at a search index that returns nothing. Watch what it does. If it retries forever, loops, or invents data to cover the silence, you don't have an agent — you have a demo that hasn't met production yet. The teams that survive their first real outage are the ones who rehearsed it on a quiet Thursday, not the ones who discovered it during a client call.
And keep a ledger. Log every tool call, its latency, and its outcome — success, timeout, error, empty. After a month you'll know which dependency is actually your weakest link, and it will rarely be the one you guessed. Reliability work is unglamorous and it never shows up in a demo, but it is the difference between an agent your team trusts and one they quietly work around.
We build AI agents that fail gracefully when the systems around them don't — timeouts, retries, and honest degradation designed in from the start — because an agent's job isn't to be right when everything works. It's to be safe when something doesn't.
Field note: A client's support agent started mis-booking refunds after a billing API began returning stale cached responses — still a 200, still valid JSON, just three hours old. The agent trusted the response and acted on it. The fix wasn't a better prompt. It was a ten-line freshness check: every billing response now carries a timestamp, and anything older than five minutes gets treated as an outage, not an answer. The agent's refund accuracy didn't change. Its worst-case behaviour did — from "confidently wrong" to "honest and stopped." That is the whole job.
Related service Workflow automation



