Giving an agent access to a tool is different from enabling it to complete professional work. An interface tells it how to call the tool; a workflow must also explain when to call it, how to judge the result and what to do after failure.
In quantum calibration, one measurement often determines the parameters of the next. Handing a model an existing notebook does not automatically capture the judgments an expert makes along the way.
Start with decision criteria
Experts know which signals are credible and which anomalies require another measurement. Capture parameter ranges, expected features and conditions that require stopping or human review. A tool can run successfully while the task still fails.
Return inspectable evidence
A success flag is rarely enough. Raw arrays, images, fitted parameters and result identifiers should point to the same operation. People need to revisit the evidence behind an agent’s decision.
Include recovery in the workflow
Retrying, changing parameters, skipping and rolling back have different consequences. Each needs a trigger and a defined scope. Preserve the state and return to an expert when a result falls outside those boundaries.
Evaluate complete tasks
Check prerequisites, persistent state, recovery and final outputs—not just isolated tool calls. Begin with a bounded workflow and compare outcomes against agreed acceptance criteria.
These observations draw on research involving members of our team. New domains require their own tools and criteria, developed with domain partners.