A 15-role AI team that builds software — governed, not unattended
Claude Corp is a multi-agent engineering system built on Anthropic's Claude. Specialized AI roles plan, build, test, and review across a portfolio of projects inside explicit decision rights and risk gates — production changes are gated by explicit release rules, all governed by a living Constitution.
The orchestration pipeline
Select a stage to follow a task from intent to merge.
A real system, with receipts
Extraordinary claims deserve evidence. Everything here is verifiable — open the live apps, inspect the source, and read how the system is governed.
Versioned governance
- v1.12 Living Constitution — Art. 17 merge gate
- OS Agent OS runtime overlay, versioned per release
- order Sky → Constitution → roles → skills
Governance is dated and citable — not vibes. No agent self-amends, and merges to main need Sky's approval — apart from the one narrow, audited exception the Art. 17 gate spells out.
Every role, its own branch
- feat/claudecorp-p1-glass-foundation
- attribution/claude-corp-truth-2026-06-18
- content/portfolio-voice-pass-2026-06-10
- perf/auto-2026-05-31
Real, role-prefixed branches across projects. Agents propose on scoped branches — production merges stay with Sky, except the single gated exception above.
This page's own 6-phase build history (opens in new tab)Continuously audited
900+
QA & audit reports written across the portfolio.
Multi-agent reviews truth-audit each other against live code and git history before anything is called done.
From backlog to review, fully automated
Each role is a Claude agent with a scoped skill file. Morgan orchestrates across roles; the Constitution enforces every safety gate, and production release is bounded by explicit governance rules.
Sky sets intent
Sky writes a directive or runs a slash command. The authority order locks in: Sky > Constitution > role files > skills.
Orchestrator plans
Morgan reads backlogs, open branches, and QA reports — then routes work to the right specialists in dependency order.
Roles execute
Each agent works in isolation on a scoped branch. Design, code, test, a11y, performance, and privacy checks all run before merge.
Safety sweep
A 7-layer Design Compiler gate plus the Orion recovery agent ensure no UI ships without full parity and a rollback path.
Sky approves release
Merges to main need Sky's sign-off, with one narrow, audited exception: a single already-shipped tool that can merge and release itself once every check passes. Everywhere else, Morgan briefs and Sky decides.
15 specialized roles
Each role is a purpose-built Claude agent with a scoped skill file, a prefixed branch namespace, and a silent external presence — only Morgan talks to Sky.
Select any role to see its authority, limits, and constraints.
Built-in guardrails, not guidelines
Claude Corp is governed by a Constitution that outranks any role or skill. These rules are enforced as governance law, not left to individual judgment.
Living Constitution
A versioned governance document (currently v1.12) that all 15 roles inherit as hard law. No agent self-amends — only Sky changes it.
7-Layer Design Compiler
Every UI change passes tokenization, accessibility parity, component consistency, visual entropy, luxury score, regression safety, and the compile decision before it's marked done.
Background Mode
Audit agents can run autonomously on a schedule, bounded by Background Mode — a halt-sentinel file freezes all background work instantly, and every cycle is capped at one reversible change.
main is governed
Every change lands on a role-prefixed branch and gets reviewed. Merging to main needs Sky's approval — except one narrow, audited exception that lets a single tool release itself once every check passes.
One voice to Sky
Morgan is the only agent allowed to message Sky. Every other role surfaces findings to QA reports; Morgan picks them up and briefs.
Velocity Loop
Five chained commands drive a continuous backlog-to-PR cycle: plan → work → review → state-sync → next. No idle time between iterations.
Built by one developer, powered by fifteen agents
Claude Corp is the multi-agent engineering system Sky Halisky designed and runs on Anthropic's Claude — always on, always scoped, always accountable to Sky. The software it maintains is real, live, and linked above.
Sky is open to thoughtful product & implementation collaborations.