The hardest part of running AI agents in production isn't the LLM
retained text
Decided to drop another discussion here since I finally found a active sub lol...yhe production question I'm most interested in is what happens after the model has already made a reasonable decision. Imagine an agent that needs to update a record, notify another system, and wait for a person. The first action succeeds. The second times out. The person responds after a deployment. None of those problems is fixed by a more articulate model. I'd want to know which actions actually completed, which results were persisted, which requests might still be running, and which approval applies to which version of the proposed action. A log that says "tool failed" doesn't answer all of that. This is why Diagrid Catalyst's focus on durable execution interests me more than another abstraction for composing prompts. Making the lifecycle a shared infrastructure concern seems useful. But I'd evaluate it by how clearly I can understand and recover a failed run, not just whether a demo restarts successfully. A few tests would matter to me: killing a worker around a tool call, receiving the same callback twice, canceling during an external request, and deploying while a run is waiting. Each needs an expected business outcome, not merely a green process-health indicator.Model quality obviously still matters. My point is that a good answer and a correctly completed task are separate outcomes. For an agent with external actions, I'd want both measured before calling it production-ready.
document fetched 2026-10-22 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xbfldb/the_hardest_part_of_running_ai_agents_in