What "production-grade" actually means for AI agents
"Production-grade" is the most overloaded phrase in the agent space right now. Every vendor uses it. Almost nobody defines it. The result is a buying process where everybody nods at the demo and the pilot quietly stalls at integration time.
Here is the working definition we use internally, five primitives that, in our experience, separate an agent that ships from one that lives in a sandbox forever.
1. Confidence thresholds are explicit and tunable
A production-grade agent does not just emit an answer. It emits an answer plus a confidence score the operations leader can act on. The threshold for auto-resolution is set per-workflow and tuned weekly. There is no "magic 95%." There is a number you and your supervisors agreed on, and a way to change it without redeploying.
2. Fallbacks are first-class
When the model returns low confidence, when an upstream API times out, when the policy engine fires a flag, the agent has a defined fallback. That fallback is not "log an error and pray." It is a routed handoff with the full context pack pre-built for the supervisor. Building this on top of a vanilla LLM SDK is a six-month project people consistently underestimate.
3. Rollback is a button, not a postmortem
Every action the agent takes is reversible. Refunds reverse cleanly. Vendor records roll back. Tickets reopen. This is not optional. The first time something goes wrong, the difference between "we hit a button" and "we are doing manual cleanup for a week" is what determines whether the rollout survives.
4. The audit trail is structured, not free-text
Each decision logs the inputs, the model output, the human's call, the policy clauses that fired, and the rationale. Structured. Queryable. Exportable. Pulling a clean trail for compliance is a SQL query, not a forensic project. The auditors who actually visit you do not want a Notion page; they want a CSV.
5. The eval harness survives a model swap
The big one. The model under your agent will change, and the change will not be your decision. A production-grade system has an eval harness, a labeled set of inputs, expected outputs, and accuracy metrics, that lets you swap models in dev, validate against the harness, and decide on the swap before it ships. If you do not have this, your "production agent" is a model-version landmine.
What it looks like when you do not have it
Most pilots that stall at the 78% accuracy ceiling are missing two or three of the five. The fix is rarely a different model. The fix is the missing scaffolding. You can read more about that bottleneck in the eval harness gap.