ENDING A RUN

STRUCTURED FAILURE REPORTS When a run fails, a vague message such as "something went wrong" wastes the next person's time. A structured failure report is a failure description in a fixed, predictable layout. A useful report contains four parts. Status says whether the run failed or partly succeeded. Reason says what went wrong, in plain words and with specifics. Partial progress says what was completed before the stop. Recommended next step says what a person or program should try next. Because the layout is fixed, programs can read it, and a different engineer can resume the work without reading the whole raw log. HANDING RESULTS BACK TO THE USER A final answer for a person should be clear, short, and honest about limits. It states what was done, what was not done, and where any files were saved. TO THE NEXT AGENT When one agent passes work to another, as you will study in Part Ten, it should hand over structured output, not a raw chat transcript. The downstream agent needs the key findings, not every search that was tried. Passing only what is needed also keeps the next agent's context small, as discussed in Chapter 18.