Agent Skills: Orchestrate

Orchestrate

Emulate the orchestrator pattern from the advisor-tool docs: a high-intelligence model plans and decomposes, then delegates the token-heavy execution to a fleet of cheap workers. Most tokens are generated at the worker rate. On BrowseComp the native version hit ~96% of Fable-solo quality at ~46% of the cost.

Target: $ARGUMENTS

This is the inverse of /advisor (cheap main loop calling an expensive advisor). Here the expensive reasoning is scarce — one planning pass, one synthesis pass — and the cheap work is abundant (N parallel workers).

Why this shape (and why a workflow)

The whole savings depend on the expensive stages actually staying thin — a planner that starts reading files, or a synthesizer that re-verifies every worker's output line by line, has silently become the expensive main loop again. A deterministic JavaScript harness (the Workflow tool) enforces that split as code: the plan phase gets a model override to the expensive tier and nothing else to do but decompose; the execute phase runs on the cheap tier by construction; there is no path for the harness to "decide" to skip the split the way a model-driven dispatch might drift under pressure. The in-instance path below is the fallback when the Workflow tool is unavailable.

Keep the expensive stages thin — do not let Plan or Synthesize do the workers' job

You may adjust the item granularity, the tier choice, and the schemas. You may NOT:

• Let the planner read the codebase to build the plan. If decomposing requires deep reading, scout the shape yourself first (Phase 0) and hand the planner a known work-list, or push the reading into a worker whose job is "read X, then report the shape of Y."
• Let the synthesizer re-do worker work. Synthesis stitches and flags conflicts/gaps; it does not re-derive results a worker already produced.
• Skip synthesis and dump raw worker output. The point of the expensive final pass is one coherent answer, not N disconnected fragments.

Phase 0: Tiers, scope, and work-list (always done by the main loop)

Do this before authoring or invoking the workflow.

1. Parse tiers and flags from the arguments:

   • --planner= — opus (Opus 4.8, $5/$25 per 1M in/out; default) or fable (Fable 5, $10/$50; reserve for genuinely hard decomposition where plan quality is the bottleneck).
   • --worker= — sonnet (Sonnet 5, $3/$15; default) or haiku (Haiku 4.5, $1/$5; only for mechanical, well-specified sub-tasks).
   • --isolation — pass if workers will mutate files in parallel and would collide; skip otherwise (worktrees are expensive).

2. Scout inline first (cheap). Before invoking the workflow, do the light discovery yourself in this session — list the files, find the call sites, scope the diff — so the plan phase operates on a known work-list instead of guessing the shape. If the work-list is already obvious from this scouting (e.g. "one item per file" over a known file set), skip the planner's decomposition entirely and pass items directly (see below) — that's an expensive planning pass you don't need to pay for.

Preferred path: dynamic workflow

When the Workflow tool is available, run the bundled template at workflow/orchestrate.js rather than hand-authoring a script per invocation:

Workflow({
  scriptPath: "<this skill dir>/workflow/orchestrate.js",
  args: {
    task:      "<the task, verbatim>",
    planner:   "opus",     // or "fable"
    worker:    "sonnet",   // or "haiku"
    isolation: false,      // true if workers collide on files
    items:     null,       // or [{ id, spec }, ...] to skip the planner and go straight to Execute
  },
})

Template pitfall: meta must be a pure literal — no string concatenation, no template interpolation, no variables in any field. The Workflow tool rejects anything else with meta must be a pure literal.

The template runs Plan → Execute → Synthesize:

• Plan (expensive tier) — skipped entirely if args.items is supplied; otherwise the planner decomposes args.task into a compact list of self-contained work items.
• Execute (cheap tier, parallel) — one worker per item, each returning a structured result.
• Synthesize (expensive tier) — stitches the worker results into one coherent answer, flagging conflicts and gaps.

It returns { items, results, answer }. Relay answer to the user; the per-item results are background detail, not the report.

Fallback path: in-instance Task dispatch (no Workflow tool)

Run the same three phases by hand with the Task tool:

1. Plan — one Task call on the planner tier: "Decompose this task into independent work items, each self-contained with no cross-item context needed. Task: <task>." Skip this call entirely if Phase 0 scouting already produced the work-list.
2. Execute — one Task call per item on the worker tier, launched together (parallel Task calls in one message, or run_in_background: true and collect notifications).
3. Synthesize — one Task call on the planner tier: hand it every worker result and ask for one coherent answer, conflicts and gaps flagged.

Guardrails

• Don't fan out what you haven't scoped. Discover the work-list first (Phase 0), then orchestrate over it. A planner guessing at structure wastes the expensive stage.
• This is opt-in scale. A workflow can spawn many agents. Use it when the task genuinely decomposes into parallel work; for a single-thread task, just do it, or use /advisor for a course-correction.
• Literal dollar amounts in this file are escaped (\$5) on purpose — $<digit> is consumed by Claude Code's positional-argument substitution and silently corrupts to an empty string or a fragment of the invocation args otherwise. See docs/advisor-and-orchestrate.md's Gotchas section.

Related

• /advisor — the inverse: a cheap main loop consulting an expensive model on demand.
• Economics, price gradient, and the native benchmarks: docs/advisor-and-orchestrate.md.