--- title: "Production Failures: Why Your AI Agents Fail Without Telling You" description: "Dispatch loops, human bottlenecks, cost leaks: the typical failure modes of AI agent systems in production, and why visibility is the only answer that holds." created: 2026-07-28T14:00:00+02:00 updated: 2026-07-28T14:00:00+02:00 tags: ["IA", "Agents", "Architecture", "Retex"] image: /images/blog/pannes-agents-ia-production-cover.webp card: /images/blog/pannes-agents-ia-production-card.webp --- An agent system in production will fail. That is not a hypothesis: it is a statistical certainty. The real question is not whether failures happen: it is whether you see them. What becomes apparent when operating a multi-agent system across several production projects simultaneously: the most expensive failures are not the loud ones. Not the 500 errors, not the timeouts that trigger an alert. The costly ones are silent: they run, they consume, they block, without ever signaling anything. Here are the pathologies that appear consistently. ## The Silent Dispatch Loop A ticket moves to "in progress." The agent fails (external reason, API quota reached). The orchestrator detects the failure and moves the ticket back to pending. The automation reassigns. The agent starts again. The agent fails again. Without a log, without a retry counter, without an explicit exit condition: this cycle runs until the quota is exhausted or until someone happens to check the dashboard. In July 2026, this scenario played out on an urgent ticket: EU AI Act compliance, a hard regulatory deadline. The monthly model quota was reached. The dispatch mechanism (automatic reassignment on failure) kept bouncing the ticket between `Todo` and `InProgress`. An agent comment documented the cause; the ticket was manually moved to `Blocked` with the explicit reason noted. The failure was dated, traced, resolved in minutes once visible. Without instrumentation, that delay would have been invisible and endless. With a structured log: a dated ticket, an identified cause, an assigned resolution. ## The Human Bottleneck This is the least anticipated category. The agent has prepared everything (market analysis, editorial strategy, format recommendations) and is waiting for fifteen minutes of human action to continue. The ticket has been blocked for 47 days. This is not an agent failure. It is a failure of visibility into what the system is waiting for from you. A ticket created on June 11, 2026, on one of our projects contains a complete plan for opening a LinkedIn page: editorial guidelines, content calendar, FR and EN bios. It is waiting for a few minutes of manual action to unblock what comes next. The orchestrator keeps this ticket visible, with its day counter running. Without that concrete representation, the delay has no form: it is a vague intention somewhere in a mental list. ## The Cost Leak Do you know what a run of your agent system costs? Not roughly: to the log line, with the timestamp and the model name. Our cost log holds over 800 entries. A complex orchestration run (multi-project analysis, dispatch, evaluation) costs around $4. That number is not surprising: it is the expected result of a correctly sized system. What would be surprising is not knowing it. Without per-run accounting, costs accumulate opaquely. A traffic spike over three days, a model swapped out accidentally, an undetected dispatch loop (see above): without the timestamped line showing the unit cost, you receive an end-of-month bill with no way to trace where it came from. ## The Ghost Deployment The automation is in place, tests pass, deployment succeeds. But the actual behavior has drifted from the intent. The agent responds, produces outputs, raises no errors, and operates on assumptions that became stale after the last prompt update. This pathology is less frequent than the three above, but often the hardest to detect. There is no failure to trace: just a silent divergence between what the agent does and what you believe it does. Instrumentation alone is not enough here: you need periodic output evaluation mechanisms, not just execution logs. --- ## What Visibility Changes These pathologies exist in every agent system, instrumented or not. An uninstrumented system encounters them just as often: it simply does not see them. The loop runs. The bottleneck accumulates. The costs go opaque. The drift is imperceptible. In a properly instrumented system, every failure becomes a ticket: dated, counted, assigned. This is not a promise of zero failures. It is the guarantee that they do not stay hidden. What an audit looks for first is not the list of bugs: it is the state of the measurement instruments. How many of your agents operate without a cost log? How many of your automations can loop without an exit condition? How many decisions are waiting on untracked human actions? An outside perspective sees what an internal team gradually normalizes. The [AI automation audit](/offres) installs the instruments and reading routines, so the next failure becomes a ticket, not a surprise.