An agent earns trust through work people can inspect.
Moving from a prompt to a real outcome takes more than fluent language. It takes planning, useful tools, feedback, deliberate permissions, and a reliable way to tell whether the result is right.
The central question.
How can agents work over long tasks without losing the original goal?
Questions that follow
- How should external actions be authorized, observed, and reversed?
- Which checkpoints give people the clearest view of progress and risk?
How we study it
A useful evaluation follows observable decisions and changes: task criteria, tool results, permission boundaries, recovery from errors, and the evidence a person needs to review the outcome.
What meaningful progress looks like
Measure successful completion, recovery after failure, adherence to permissions, and the human effort needed to review or correct the result.
Explore all research programsRead the work behind these questions.
These independent papers and official guidance offer useful starting points for this field. Follow each link to review the original methods, evidence, and limits in context.