An agent closes a support ticket. The answer it gives is wrong. The customer writes back, angrier. A person takes the ticket, fixes it in ten minutes, and the postmortem gets one line. The model hallucinated.
The fix ships the following week. A bigger model, a newer version, a different vendor. The ticket count looks the same. The wrong answers look the same. They are written better now.
A second team, a second agent. This one approves a refund it should have declined. The logs are good. They show the prompt that went in and the answer that came out. They do not show the note on the account from the week before. They do not show the policy that changed on Tuesday. Nothing in the logs shows what the agent was never handed, because nothing records an absence.
This is not two teams. This is the shape of the postmortem for most agent failures that will happen next year.
Most of those failures will be context failures. People will blame the model for nearly all of them.
The model gets blamed because it’s the only part of the system with a name. It has a version number. It has a price per token. It has a vendor, a status page, and a sales rep. When something breaks, the model is the one piece you can point at, cost out, and replace by Friday.
Context has none of that. Context is the account note, the Tuesday policy, the field someone stopped syncing in March. There is no vendor for it. There is no line item. Nobody owns it, so nobody gets paged when it is missing.
A model failure is a purchasing decision. A context failure is someone’s job.
The logs make it worse. Logging was built to record what a system did. It records the prompt and the response. It doesn’t record what the agent should have been shown but wasn’t, because the system that assembled the prompt didn’t know either. The gap never reaches the log. The guess does.
Swap the model and the guess gets more fluent. The gap stays exactly where it was.
Look at what gets measured today. Eval scores. Benchmark rank. Token cost. Latency. All of it is about the model. None of it is about the inputs. No team I know has a number for how stale the agent’s picture of the world was when it acted. Nobody reports what fraction of the relevant facts made it into the prompt. That number exists. It is just not on a dashboard, so it is not anyone’s problem.
The person who did this job before the agent knew about the Tuesday policy because they were in the Tuesday meeting. They knew about the account note because they wrote it. They carried the context in their head, and nobody ever had to log it, because the person and the context were in the same place.
The agent was not in the meeting.
The upgraded model gets the same ticket next month. It reads the same prompt. It answers with more confidence and better grammar. The account note is still missing.