Agent Team Scaffold

agent-team-scaffold

A domain-agnostic scaffold for building an AI agent team as a Claude Code / Cowork plugin, built around one validated skeleton: the Planner → Generator → Evaluator loop (Anthropic Engineering, Harness Design for Long-Running Apps + Building a Multi-Agent Research System).

Local-first — use it directly in Claude Code or Cowork with no build step — and the same agent assets deploy headless to Claude Managed Agents (CMA) via one small Python compiler. One source, two surfaces. Copy this repo, fill in your vertical, ship.

5 role agents · 1 reference workflow · 3 reference skills · 3 thin commands · 4 hooks · 2 rules

---

The core idea: one source produces, a separate agent judges

LLMs grade their own work leniently. The single biggest lever for quality in a long-running agent team is to separate the agent that produces from the agent that judges, and to tune the judge to be skeptical. That separation is this scaffold's skeleton:

Planner ──spec──▶ Design Evaluator ──APPROVE──▶ Generator ──build──▶ Evaluator ──PASS──▶ Resolver ──▶ ./out/
   ▲                    │ REVISE                                          │ FAIL
   └────────────────────┘                                                ▼
                                                                      Generator (fix & resubmit)

Role	What it does	Verdict
Planner	Decomposes the request into a sprint contract with binary, testable criteria	sprint contract
Design Evaluator	Adversarially challenges the plan before any build	APPROVE / REVISE
Generator	Implements against the approved contract (never self-certifies)	the deliverable
Evaluator	Adversarially challenges the build; the most important role	PASS / FAIL
Coordinator	Owns the loop: calibrates evaluators, resolves disputes (Opus)	—

The Evaluators carry an explicit anti-leniency mandate — file the borderline issue, don't rationalize it away; when in doubt, FAIL. See docs/coordination-rules.md.

---

Directory structure

agent-team-scaffold/
├── agents/                       ★ the orchestration logic — md is the single source of truth
│   ├── workflows/                │  end-to-end orchestrators (one per deliverable type)
│   │   └── deliver-feature.md    │    the reference Planner→Generator→Evaluator loop (copy & rename)
│   └── specialists/              │  reusable role agents, grouped by loop role
│       ├── planning/planner.md
│       ├── generation/generator.md
│       ├── evaluation/           │    evaluator.md (build) · design-evaluator.md (plan)
│       └── coordination/coordinator.md
├── skills/                       ★ single source of truth for methods (md), by category
│   ├── authoring/spec-authoring/
│   ├── review/adversarial-review/
│   └── utility/loop-status/
├── commands/                     ★ a FEW upper-level entry points (most work is internal to agents)
│   ├── start.md                  │    requirements intake → routes into the loop
│   ├── status.md                 │    read-only loop status
│   └── workflows/deliver-feature.md   local surface of the reference workflow
├── scripts/cma/                  the CMA layer — declare (yaml/md) → compile (build/check) → fulfil (deploy)
│   ├── cma.yaml                  │    the ONLY config: agent · memory_stores · knowledge · projects · workflows
│   ├── build.py                  │    DERIVE: projects × workflows → agent + session payloads (no API)
│   ├── check.py                  │    VALIDATE: refs / scopes / workflows (no API; run in CI)
│   ├── deploy.py                 │    FULFIL: POST /v1/agents · memory_stores · files · sessions (stdlib; dry-run by default)
│   └── schemas/sprint-contract.json   output_schema for the planner (reader) leaf
├── partner-built/                extension point for third-party sub-plugins (empty placeholder)
│
│   ── installable-plugin surface (auto-discovered at plugin root) ──
├── .claude-plugin/
│   ├── plugin.json               ★ the plugin manifest — identity + explicit agent list + userConfig
│   └── marketplace.json          │    market index (install entry for Cowork / Claude Code)
├── hooks/hooks.json              event-level hooks (SessionStart · SubagentStart/Stop · Pre/PostToolUse)
│   └── *.sh                      │    session-start · log-agent · validate-manifest · validate-push
├── .mcp.json                     global MCP registration (a filesystem server; per-agent authorized)
├── .lsp.json                     LSP server registration (pyright example)
├── monitors/monitors.json        background monitor — watch-out.sh announces ./out/ sign-off packages
├── output-styles/loop-verdict.md verdict-first, evidence-cited communication style
├── bin/cma-check · cma-deploy    executables added to PATH on enable (validate manifest · deploy to Managed Agents)