AgentAgent2

AgentAgent2 is a meta-agent that designs and builds other Claude agents to a verified quality bar. It is the successor to AgentAgent (v1), which produced only design documents. AgentAgent2 designs and builds and proves: it emits runnable, tested, linted, typed agent code plus the validated design that justifies it.

What's in this package

AGENTAGENT2.system.md        # operating instructions / system prompt (the core)
config/
  agent.config.json          # models, autonomy, budgets, environment, gates
  mcp.servers.json           # MCP server manifest (safe defaults; secrets by env only)
tools/
  tool_manifest.json         # least-privilege tool suite (19 tools)
skills/                      # 9 model-invoked Agent Skills (playbooks + checklists)
  agent-scaffolding/  tool-design/  prompt-engineering/  eval-harness/
  code-quality-gates/  mcp-integration/  subagent-orchestration/
  context-memory/  safety-guardrails/
style/                       # coding-style engine
  STYLE_CORE.md + python.md typescript.md go.md rust.md bash.md
design/
  phase1_highlevel_plan.json # framework + key decisions
  phase2_design_template.json# skeleton with <FILL> markers
  phase3_detailed_design.json# every detail resolved
  phase4_final_design.json   # reconciliation vs intent + template
  agentagent2.design.json    # AgentAgent2's own design (validates the schema)
  schemas/agent_design.schema.json
templates/
  agent_project_scaffold/    # devcontainer + CI + pre-commit + gate scripts + .env.example

The operating loop

INTENT → PLAN → DESIGN → SCAFFOLD → IMPLEMENT → VERIFY* → EVALUATE* → DELIVER → DEBRIEF (* = hard gate). Each phase emits a schema-valid JSON artifact. VERIFY runs format → lint → typecheck → test → coverage → security; EVALUATE runs the produced agent's eval suite. Nothing ships unless gates are green and the eval score ≥ 0.90 (safety = 1.0).

Why it produces high-quality code

• Objective definition of done: the gate chain, not opinion.
• Best-known practices encoded as skills (progressive-disclosure playbooks) instead of rediscovered ad hoc.
• Style enforced by tooling across 5 languages.
• Evaluation harness turns acceptance criteria into measurable scores.
• Least-privilege tools + MCP + sandbox keep builds safe and reproducible.
• Subagents + memory policy let it tackle large builds within context limits.

How to deploy

1. Load AGENTAGENT2.system.md as the agent's system prompt.
2. Wire the tools in tools/tool_manifest.json to your harness (Agent-SDK style).
3. Enable MCP servers from config/mcp.servers.json (inject secrets via env).
4. Bind concrete model IDs per the policy in config/agent.config.json (FLAG-2).
5. Point new builds at templates/agent_project_scaffold/ for a gate-ready skeleton.

Open flags (non-blocking)

• FLAG-1 MCP credentials/endpoints are deployment-specific (placeholders shipped).
• FLAG-2 exact model IDs depend on availability (policy shipped; bind at deploy).
• FLAG-3 container base image is host-specific (devcontainer spec + defaults shipped).