A.R.C.AAdversarial multi-agent system · ML / AI engineering
The adversarial agentic system that actually ships code.
59 specialized agents. 3 pipelines. 111 ADRs. Zero AI slop tolerated.
What A.R.C.A does
Seven pillars that turn a single AI coding assistant into a regulated multi-agent engineering team. Pure configuration — no runtime, no server, no container.
59 Specialized Agents
Each with its own model assignment (Opus for reasoning, Sonnet for implementation), specific role, blocking gate and audit responsibility. From @data-validator to @ai-redteam-orchestrator.
Adversarial Gate Chain
Producer → @math-critic → @debt-detector → @code-critic → @chief-architect. No code reaches main without explicit sign-off. 65 bash hooks enforce the chain at PreToolUse / PostToolUse level.
3 Pipelines (ML · HTB · ART)
ML: C1→C14 (14 cycles, 65 phases). HTB: F0→F7 (CVE-first CTF with MCP integration). ART: R0→R8 (AI Red Teaming, MITRE ATLAS + OWASP LLM Top 10:2025, ADR-081).
111 Documented Decisions
Every architectural choice ships as a Nygard ADR: context, alternatives weighed, consequences. Covers architecture, security, governance, pipelines and the meta-system itself.
148-Skill Catalog
OWASP security, ML engineering, DevOps, RAG patterns, agent orchestration, HTB CTF methodology, AI red teaming, prompt engineering. @skill-router selects ≤3 per task.
Honest Disclosure
Inverted-style canary tests assert known bugs still occur, so a future change cannot silently mask a structural limit. 19 AI-slop signals detected pre-merge. Zero hidden state.
Dynamic Orchestration
A 4th mode for the long tail no fixed pipeline covers. @architect-ai proposes a bespoke agent DAG — which subagents, in what order, with which adversarial critics per node — for Adrian's approval before any agent executes. Two-layer gate enforcement: a schema-floor validator rejects under-gated proposals at declaration, a runtime hook hard-blocks at invocation. /orchestrate · ADR-089.
Quick look
A.R.C.A lives inside Claude Code’s native extension points: markdown agent definitions, bash hooks, JSON settings. Clone the repo, run install, work.
# 1. clone
$ git clone https://github.com/infantesromeroadrian/arca-claude-code
$ cd arca-claude-code
# 2. install — symlinks settings + hooks + skills into ~/.claude
$ ./install.sh --skills --hooks --mcp
# 3. verify the gate chain
$ claude --debug --agent code-critic 'review this'
↳ @math-critic
↳ @debt-detector
↳ @code-critic
↳ @chief-architectThe gate chain
Four adversarial agents stand between producer code and main. Each one blocks. Bypass leaves an audit trail.
@math-critic
Validates loss functions, gradients, numerical stability, attention scaling, sampling strategies. Blocking on producer code from @ml-engineer / @dl-engineer / @ai-engineer in C3, C5, C6, C8.
@debt-detector
Catches unused imports / functions, TODOs without tickets, logical duplication (>3 sites), cyclomatic complexity over 10. Blocking on producer output in C6 and C8.
@code-critic
19 isolated AI-slop signals. Audits architecture, strict security adherence, runtime performance, deep maintainability limits. No code is final without explicit approval.
@chief-architect
C10 deploy gate. Cross-checks that @code-critic, @math-critic, @model-evaluator, @tester and @ai-red-teamer signed off. Sin firma, no hay deploy.
Built on
The stack underneath. Each one earned its place with a documented ADR.
The 14-cycle pipeline
From Discovery to Sunset across 65 phases. Each cycle has a blocking gate at its exit — owned by a specific agent.
ML, HTB and ART are the three fixed pipelines. For the long tail no fixed pipeline covers, @architect-ai proposes a bespoke agent DAG — chosen subagents, their order, and the adversarial critics gating each node — for Adrian’s approval before any agent executes. A schema-floor validator rejects under-gated proposals at declaration; a runtime hook hard-blocks at invocation. ADR-089.
Architecture decisions
111 numbered ADRs. Every architectural choice carries context, alternatives weighed and consequences. Below: 19 milestone records.
The 9 mortal sins
Hardcoded violations that block the cycle and trigger visible escalation. No warnings — direct block.
Stack
Runtime posture, model distribution, persistence, compute. No marketing — just what is actually wired.
- MODELS
- Opus 4.8 (1M ctx) · Sonnet 4.6 · Haiku 4.5
- DISTRIBUTION
- 51 Opus · 4 Sonnet · 4 Haiku (59 agents · 3 fixed pipelines ML + HTB + ART + 1 dynamic orchestration mode)
- ENTERPRISE TIER
- 16 agents v3.0 regulated-grade (SOC 2 / EU AI Act / GDPR / DORA / HIPAA / PCI-DSS)
- PERSISTENCE
- Engram MCP (Local SQLite)
- LLM-AS-JUDGE
- Hybrid: Opus 4.8 SDK (high-stakes) + Qwen 2.5 7B (hot-path)
- RUNTIME
- 100% bash hooks (Zero startup latency)
- PLAN TIER
- Claude MAX flat-rate
- COMPUTE NODE
- RTX 2000 Ada Lovelace SM 8.9 (8GB VRAM)
Take it for a spin
A.R.C.A is open source. Clone it, audit it, fork it. Built for the Anthropic Fellows 2026 application.
Powered by