Agents fail in production for ordinary reasons: unclear goals, missing stop conditions, and no human review when confidence is low. Scaling those failures makes them more expensive.
Codify the work before you automate it
If a team cannot write the steps a skilled person follows, an agent will improvise. Capture the happy path, the exceptions, and the decisions that require a person. Then decide which steps are safe to automate.
Guardrails are part of the product
- Limit tools, data access, and write permissions by default.
- Require a human checkpoint for irreversible actions.
- Log prompts, tool calls, and outcomes so failures can be inspected.
- Test adversarial and ambiguous instructions before a wide rollout.
Culture has to match the architecture
If operators are rewarded only for speed, they will bypass review. If they are rewarded for catching bad agent behavior, the system improves. Scaling sticks when the workflow and the incentives say the same thing.
Start with a narrow loop
Run one agent on one workflow with a measured evaluation set. Expand only when error types are known and the review path is staffed. That is slower at the start and cheaper later.