dynamic-workflows

Dynamic workflows

Deterministic multi-agent orchestration via the built-in Workflow tool — a JavaScript script that fans work out across subagents (agent()), pipelines it (pipeline()), or barriers on it (parallel()). This skill decides when orchestration earns its cost and gives the minimal scaffold to launch one. The Workflow tool's own description carries the full API — lean on it; don't reproduce it here.

Mode detection (FIRST step every time)

Check the system prompt for the marker Dynamic-workflows mode (loop).

• Marker present → LOOP mode. You are inside claudenv loop. Orchestrate without pausing, but only for plan items that genuinely split into independent sub-tasks, and keep fan-out narrow (see Cost discipline). The goal is law — never pause to ask "запустить workflow?".
• Marker absent → INTERACTIVE mode. Propose a one-line orchestration plan (how many agents, what each does, pipeline vs parallel) and wait for the user before launching.

When to orchestrate (triggers)

Reach for a workflow when the work is wide — many items, each handled the same way, with little cross-talk:

• Review / audit a diff across several dimensions (bugs, perf, security, style), then adversarially verify each finding.
• Migrate / edit a pattern across many files (worktree isolation per agent when they mutate in parallel).
• Research a question over many sources, then synthesize.
• Generate N independent attempts (different angles), judge, synthesize from the winner.
• "Find all X" with a verification pass and loop-until-dry.

When NOT to orchestrate (anti-triggers)

Stay single-threaded — a workflow is pure overhead here:

• A single file, or a change that must happen in strict sequence.
• Trivial edits, renames, formatting.
• Anything one or two plain Agent calls already cover — prefer those.

Cost discipline. Multi-agent fan-out costs roughly an order of magnitude more tokens than single-threaded work (~15×; see claudenv-update-plan.md). Single-threaded is the default. Orchestrate only when the width is real, and cap the number of concurrent agents to what the task needs.