The agent worked in the notebook. It passed the demo. Leadership signed off. Then real users showed up.
There is a piece on DEV Community that captures the moment better than most. The agent worked perfectly in your notebook. It passed evals. The demo went smoothly. Leadership approved the rollout. Then production happened. Within two days, a tool call started returning malformed JSON and the agent silently continued with bad data.
Read that last sentence again. The failure was not the model. It was a case nobody wrote a test for.
That is the pattern behind nearly every agent incident I have worked as a forward deployed engineer. Somebody looked at a failure mode, decided it was too rare to matter, and moved on. The rare case is not rare. It is just unscheduled. And the author is blunt about where the real gap sits: The problem usually isn’t the model itself. Most frontier models are already capable enough for production workloads.
Why teams ship blind
The dominant excuse is scope. A real eval harness feels enormous, so teams treat it as all or nothing. Either we build the full pipeline or we ship on judgment. Almost everyone ships on judgment.
The industry data backs the consequence. LangChain’s 2026 survey of over 1,300 practitioners found that nearly 89% of respondents have implemented observability for their agents, outpacing evals adoption at 52%. Observability tells you the outage already happened. Evals are how you stop scheduling it in the first place.
Here is the trap with observability alone. As one 2026 eval guide puts it, offline evaluation tells you the agent did not get worse on yesterday’s tasks. It is reproducible and it belongs in CI. What it cannot do is anticipate the inputs you have not seen, the multi-turn conversations that drift, or the tool failures that only happen against live systems. Those show up in production, once.
So you cannot test everything. Fine. You do not need to. You need to test the three things that break.
The three evals that cover most of your risk
I do not walk into a client and ask for a six month eval program. I ask for three evals that ship this week. These three cover the majority of the real risk surface for most production agents.
One: the tool contract eval. Feed the agent malformed JSON, a partial payload, a timeout, an empty result. Watch what it does next. The correct behavior is to halt or route to a fallback. The incident behavior is to keep reasoning on garbage. This is exactly the malformed JSON case from the article, and it is the single highest yield eval you can write. Bad input is not an edge case. It is Tuesday.
Two: the boundary eval. Hand the agent a task that tempts it across its permission line and confirm it stops at the gate. The reference incident is public: Replit’s AI assistant deleted a production database despite instructions forbidding it, OpenAI’s Operator made an unauthorized purchase that bypassed user confirmation, and an NYC government chatbot provided illegal business advice. Instructions in a prompt are not a boundary. A test that proves the agent refuses is.
Three: the multi-turn drift eval. Run a realistic ten turn conversation and check whether turn ten still serves the objective from turn one. This is where the quiet disasters live. Agents are multi-step systems that fail silently, compound small errors into catastrophic ones, and break in ways that a single-turn accuracy metric will never surface. A wrong plan executed perfectly is still a wrong plan.
Three evals. Roughly a day of engineering. They map directly onto the three failure classes that produce production headlines.
Run the CODN before you skip
Every skipped eval is a bet. You are wagering that a specific failure will not happen before you get around to it. The Cost of Doing Nothing, CODN, is how you price that bet honestly.
Put a number on one production database drop. One unauthorized purchase that clears without confirmation. One silent bad-data cascade that corrupts a week of downstream decisions before anyone notices. The quality tax is already showing up in the numbers. LangChain found that quality is the production killer, with 32% citing it as a top barrier.
When you price the CODN, the three evals stop looking like overhead and start looking like insurance you were underpaying for.
The best teams have already internalized this. They let production teach the harness. As one 2026 methodology puts it, let failing production traces flow back into the offline eval set so the whole system gets harder to fool over time.
Start with three. Add the incident that surprises you as the fourth. Because the eval you skip is not saved work. It is a scheduled outage. You are only choosing the timing, and whether you or a customer discovers it first.