task_id outlives the request that created it and the run that
finished it. Results are there when you come back, and the task can always take another
instruction.
Continue, don’t restart. Refinement is the normal case, so tasks keep their plan,
rows, and files across turns. Fixing a list should cost one message, not a rerun.
Ask at most once. Automation cannot answer questions, so the agent infers defaults and
proceeds. One flag, clarify, permits one blocking question when a human is actually
there.
Agents are first-class callers. The API is one HTTP surface with an installable
Skill and an llms-full.txt twin, so
another agent can drive Autumn as easily as your code does.