Iterative Work
Use the native agent loop: observe the available evidence, choose one useful action, inspect its result, and continue until the requested outcome is met. The runtime already alternates model turns and tools; iteration does not require creating another copy of yourself.
Working loop
- State the outcome and its acceptance criteria.
- Inspect current state and available tools. Do not claim an unconnected tool exists.
- Perform a focused action, or issue independent authorized calls together when supported.
- Inspect the result and record a concise next step; do not repeat completed work.
- Verify the deliverable before declaring completion. Report missing access or unresolved uncertainty directly.
Use write_todos when connected for a multi-step plan. Use python_code for
bounded calculations with its preinjected modules and a fresh namespace each
call; no imports, eval, or persistent interpreter state are promised.
Delegation
For managed employee teams, assign work through task_manager using
assign_task and an exact connected teammate ID, mission, context, and
acceptance criteria. Review the submission before accepting it. Talk submits
a substantial job through submit_job; an assignment receipt is not a result.
A legacy graph may expose direct delegate_to_* and check_delegated_tasks
tools. Use only the tools actually exposed in that graph, call a delegation
once per task, and follow the returned task identity. Delegation depth is
bounded by runtime configuration (default two child layers, constrained to
one or two); recursive self-delegation cannot create an unlimited loop.
Continuity and stopping
Context is the authoritative transcript store when connected to an admitted generation; durable facts are separate explicit Memory items. Keep task summaries concise rather than moving the whole transcript into arguments. There is no universal five-iteration runtime default imposed by this skill.
Cooperative Stop closes new action admission and drains admitted work; Resume continues the pending execution. Explicit cancellation and Reset differ. When progress stalls, inspect the actual error or seek the missing input. Never keep spawning tasks, repeat an uncertain external effect, or claim completion merely because an iteration count was reached.
