AI agents will retry. They will always retry. Given how long-running agentic loops are, network drops, timeouts, machines rotate, and rate limits kicking in are just part of the deal.
What breaks in production is not the model, but the plumbing around it. Some gaps worth noting…
First - A half-dead (slow and stalled) API can hang an entire agent chain. So, every tool call should have a hard timeout, and when it fires. This way, the agent fails loudly instead of waiting indefinitely.
Second - A retry here and there is fine (and essential). But if a downstream tool is actually down and you keep hammering it, you burn through your quota for nothing. So, have a backoff on failure, then trip the circuit after a run of n consecutive failures, then probe again later. Do not retry blindly into a dead service.
Third - If the process dies halfway through a long task, can it pick up from where it stopped, or does it start over from scratch? That only works if each step’s progress is written somewhere durable, not just kept in memory.
Fourth - Tracing. Once an agent is chaining multiple steps, “something felt slow” is not debuggable. You need to see exactly which step took eight seconds instead of two.
None of these four are hard on their own. We have all done this while designing traditional systems, and those practices are not going anywhere :) We just need to pack them together in one system.
That is really the bar for production readiness. You do not need a smarter model; you need a rock-solid harness and plumbing around your agentic loop. Everyone loves a reliable system.
Hope this helps.