dynamic-workflow

Move the plan and loop into a script instead of the agent context. Run a deterministic pre-pass with execute_code to build a manifest, then issue bounded delegate_task batches for the LLM-judgment steps, keeping only the final verified answer in context.

Fan out the same question to 2-4 independent agents with different framings, collect their atomic claims, then run a refuter batch tasked with breaking each claim. Surface only claims that survive refutation, and iterate up to 3 rounds until no new claims emerge.

No, delegate_task runs synchronously inside the parent turn and all in-flight children are cancelled if the user sends a new message or stops the turn. For durable workflows that survive interruption, use the SQLite-backed kanban swarm, a cronjob, or a background terminal process.

delegate_task is not in the SANDBOX_ALLOWED_TOOLS set available to execute_code scripts, so the import will fail. Use execute_code only for deterministic work like fetching, parsing, and templating, and issue delegate_task batches from the parent turn.

Avoid fan-out for small bounded tasks under roughly 10 units, where a direct tool call is cheaper, and for serial tasks where one unit depends on another's output. Also avoid it when the work must survive user interruption, since synchronous delegation is turn-scoped.

Each delegate_task child is a full agent tree, so 20 children means roughly 20 times the model calls. Start on a scoped slice, report the token cost, and only scale up after the user approves; do deterministic work in scripts with no per-item LLM calls.