Claude Code 32-Agent Team

A production-grade multi-agent engineering team for Claude Code, with hook-enforced protocol, adversarial review, Bayesian trust calibration, dynamic specialist hiring, an optional parallel cognitive layer, and a self-improving meta-cognitive loop.

Architecture diagrams: ARCHITECTURE_DIAGRAMS.md (editable Mermaid source + ASCII fallbacks + PNG export instructions)

This is not a framework. It's an operating system layer on top of Claude Code — 32 domain-expert agents coordinated through a syscall protocol (NEXUS), governed by runtime hooks, validated by a 352-assertion contract test suite, calibrated by a per-agent trust ledger, and equipped with a gated hiring pipeline for growing the roster as coverage gaps appear.

What it actually is (verified against itself):

• 32 agent definitions in agents/ (each 20–56 KB of role-specific instructions, spanning Go / Python / TypeScript / Elixir / BEAM / K8s / GCP / security / QA / AI platform / external BEAM retainer / recruiting + team-evolution roles)
• Runtime hooks that enforce closing-protocol invariants, persist signals, auto-log syscalls, auto-record trust verdicts, and gate post-hire registration
• A signal bus (agent-memory/signal-bus/) that carries memory/evolution/cross-agent flags between sessions
• A trust ledger (agent-memory/trust-ledger/) that Bayesian-blends per-agent accuracy from validator verdicts
• A documented NEXUS syscall protocol for privileged operations (SPAWN, SCALE, RELOAD, MCP, ASK, CRON, INTUIT, etc.)
• A Dynamic Domain Expert Acquisition pipeline (talent-scout + recruiter) — the team detects its own coverage gaps (5-signal confidence score + session-sentinel co-sign) and hires new specialist agents through 8 gated phases (research → synthesis → validation → challenger review → atomic registration → post-hire verify → probation → promotion)
• An optional Shadow Mind cognitive layer (intuition-oracle + Observer + Pattern Computer + Speculator + Dreamer + Pattern Library) that offers probabilistic pattern-based guidance via [NEXUS:INTUIT], with topic cluster pattern retrieval ("have we seen this before? what fixed it?"), delete-to-disable without affecting the conscious team
• Execution modes (FAST / BALANCED / FULL_POWER) for speed/cost/quality trade-offs per task
• 6-state agent lifecycle (candidate → probationary → active → trusted → deprecated → retired) with trust-weighted routing preference
• Quality gates by task type — different workflows for code changes vs security findings vs architecture decisions vs incidents
• Conflict arbitration protocol — structured evidence-based dispute resolution when agents disagree
• nexus-doctor.sh — diagnostic health check for framework integrity (agents, hooks, dependencies, Shadow Mind status)
• Planning cache (playbooks) — reusable task workflows that make repeated task types faster over time
• Contract tests (tests/agents/) that run on every commit and gate prompt evolution

---

Why this exists

Most agentic frameworks try to enforce discipline inside the agent loop (via code wrapping the model). This team enforces it below the agents (via hooks at the Claude Code runtime layer). That's structurally better: agents cannot talk their way past a SubagentStop hook the way they can talk their way past a Python try/except.