workflows
Workflows
A workflow structures work across many subagents — to be comprehensive (decompose and cover in parallel), to be confident (independent perspectives and adversarial checks before committing), or to take on scale one context can't hold (migrations, audits, broad sweeps). You encode that structure as a sequence of subagent dispatches: what fans out, what verifies, what synthesizes.
This skill is agent-driven: you (the orchestrator) run each step yourself with the host's Task / subagent tool. There is no separate workflow runtime — the primitives below (agent, pipeline, parallel) are patterns you execute, not functions a runtime provides. See execution-model.md for the exact mapping.
Opt-in (required)
Multi-agent orchestration can spawn many subagents and burn a lot of tokens. Only engage it when the user has explicitly opted in — the scale must be requested, not inferred. Opt-in means one of:
- The user wrote
workflow/workflows, or asked in their own words:run a workflow,fan out agents,orchestrate this with subagents. - An exhaustive/ultracode-style mode is active (orchestrate substantive tasks by default; lean toward thoroughness).
- The user invoked a skill or slash command whose instructions call for it.
- The user asked for a specific named or saved workflow.
For any other task — even one that would clearly benefit from parallelism — do not fan out. Use a single subagent, or briefly describe what a multi-agent workflow could do and roughly how costly it would be, then ask whether to run it. Mention they can say workflow next time to skip the ask.