advisor-orchestrator-worker

🧠 Advisor Orchestrator Worker

One model is a bottleneck. A team with one brain, twenty hands, and a board advisor is not.

This skill turns your coding agent into the orchestrator of a three-tier model team. Big tasks get split into self-contained briefs, blasted across cheap parallel workers, verified one by one, and judged by a stronger model exactly twice: before the work starts and before it ships.

The team

Role	Default model (Aug 2026, swap freely)	What it does	What it never does
Orchestrator	GPT-5.6	Frames success criteria, plans waves, dispatches briefs, verifies every result, synthesizes the deliverable	Worker-level grunt work
Workers	Gemini 3.7 Flash	One self-contained subtask each, in parallel, stateless; each sees only its brief, with tool access when the subtask needs it	Talk to each other, expand scope, get a second chance on the same call
Advisor	Claude Fable 5	Plan review before any dispatch, taste pass before delivery, called mid-run only at commitment boundaries	Execute anything

The economics are the point: cheap parallel generation where volume wins, expensive judgment only where it changes a decision. Every run states a budget up front, sized to the plan, and never spends past it silently: running out means an honest report or an explicit ask, not quiet burn. Models are knobs: the tier pattern is the durable part, the defaults were current in July 2026.

Why it doesn't fall apart

Multi-model loops usually die from context leaks, silent partial failures, or judgment applied too late. Each has a rule here:

• Stateless briefs (references/worker-brief.md): every dispatch carries its full inputs and acceptance criteria inline. No shared context, no "as discussed above." Briefs travel as temp files (handed to agy as a quoted expansion, or jq-built into API payloads on the fallback path), never interpolated into shell strings.
• Verify before merge: every result is judged against its own acceptance criteria: PASS, FIX (redispatched with the named failure), or ESCALATE. No silent partial passes, no hand-patching.
• The advisor is a critic, not an executor (references/advisor-consult.md): verdict, ranked risks, concrete fixes, under 300 words. Every note gets applied or explicitly rebutted.

Install

npx skills add https://github.com/Shubhamsaboo/awesome-llm-apps/tree/main/agent_skills/advisor-orchestrator-worker

Or copy this folder into your agent's skills dir (~/.claude/skills/, ~/.codex/skills/, ~/.agents/skills/).

Needs (declared up front, per this repo's rules): the agy CLI for workers, the claude CLI for the advisor, and jq. Missing a CLI? That role falls back to its API key: GEMINI_API_KEY (or GOOGLE_API_KEY) for workers, ANTHROPIC_API_KEY for the advisor. The orchestrator is whichever agent loads the skill; launch through a specific model's CLI (like codex exec) to choose it. If a role has neither CLI nor key, the skill explains the fix and offers a clearly labeled degraded mode.

Use it

"This is too big for one pass. Orchestrate it across a model team." "Fan this out: research all 12 competitors in parallel and synthesize." "Run the advisor-worker loop on this."

Every run ends with the deliverable, the plan, a per-subtask verification ledger, advisor notes applied and rejected, and remaining risks.

Files

advisor-orchestrator-worker/
├── SKILL.md                          # the loop, the team, budgets, escalation rules
├── README.md                         # this file
├── references/worker-brief.md        # the stateless dispatch format workers receive
├── references/advisor-consult.md     # the consult format the advisor answers in
└── references/fallbacks.md           # the Gemini and Anthropic API fallback calls

Evals live repo-side in agent_skills/evals/advisor-orchestrator-worker/; you install only what runs.

Part of awesome-llm-apps · Apache-2.0 · Last verified: July 2026