Task agents

Every task agent run started with runTask is recorded for 90 days. These two reads show what actually happened.

tasks list pages runs, newest started first, without transcripts. Each row's outcome says how the run ended: answered, in_progress, stopped (it asked for tools and nothing followed), or error (a model call failed). toolErrorCount and hasTurnError mark the runs worth opening. requestId names the method execution that started the task, and devSession separates your own dev runs from real traffic.

tasks get opens one run in full by its traceId, the value runTask returns. You get the task's input, each turn's reply with its stop reason and any reasoning, and every tool call with its arguments and result. A method tool's result carries a request id, so that call's input, output, console output, and error are one requests get away.

To iterate on a task agent, read the transcript, then change the prompt and the tool descriptions based on what it shows.