Home

AssemblyZero

Multi-Agent Orchestration Platform for Enterprise AI Development

Run 12+ AI agents concurrently. One identity. Full governance. Measurable ROI.

graph TD
    subgraph Intent["HUMAN ORCHESTRATOR"]
        O["Intent<br/>& Oversight"]
    end

    subgraph Agents["CLAUDE AGENTS (12+)"]
        A["Feature | Bug Fix<br/>Docs | Review"]
    end

    subgraph Verify["GEMINI VERIFICATION"]
        G["LLD Review | Code Review<br/>Security | Quality"]
    end

    subgraph Gov["GOVERNANCE GATES"]
        M["Requirements | Implementation<br/>Reports | Audit Trail"]
    end

    subgraph Orch["END-TO-END ORCHESTRATION"]
        R["LangGraph Pipeline<br/>Triage → LLD → Spec → TDD → PR"]
    end

    O --> A
    A --> G
    G --> Gov
    Gov --> R
    R -.->|"Feedback Loop"| A

Loading

---

Start Here

You are...	Start with	Then explore
Engineering Leader evaluating AI adoption	Why AssemblyZero?	Measuring Productivity · Security & Compliance
AI Strategy / Operations implementing Claude Enterprise	AI Strategy & Operations	Prompt Economics · Cerberus PR Governance
Technical Architect designing agent infrastructure	Technical Architecture · Multi-Agent Orchestration	LangGraph Evolution · Worktree Isolation
Security & Compliance approving AI tooling	Secret Guard Architecture · Fleet Security Audit	Governance Gates · Agent Instruction Hierarchy
Practitioner building with Claude Code	Quick Start · The Pipeline	The WinError 206 Hack · Permission Friction

---

Production Evidence: 936 Issues Closed, 1,273 Commits, 645 PRs

This isn't theoretical. AssemblyZero has been under continuous development for 184 days since January 10, 2026 — with commits landing to main on 129 of those days and near-daily issue and PR activity throughout.

Metric	Value
Issues closed	936
Commits	1,273
PRs merged	645
Tests	10,400+ across 290 files
Active days	129 of 184 (commits to main)
Peak day	60 issues + 89 commits + 41 PRs (Feb 3–4)

Daily Activity Log → · Metrics Dashboard → · March 2026 Velocity →

---

What's New — Q2 2026 (April–June)

398 issues closed, 321 PRs merged, and 321 commits since the March update. The platform went from a working pipeline to a fleet-scale governed system. Activity by month:

Month	Issues	PRs	Commits	Arc
April	63	61	61	Secret handling & fleet plumbing
May	224	176	183	Orchestration + governance at fleet scale (peak month)
June	111	84	77	Governance-client migration + orchestrator lifecycle

April — Secret Handling & Fleet Plumbing

A hardening month: the security patterns and fleet tooling the rest of the quarter builds on.

• In-process classic-PAT decryption (ADR-0216) — elevated-scope GitHub writes with the token living only in the Python heap, never in an environment variable or a process argument.
• pr-sentinel / Cerberus fleet standardization — every repo migrated to the fleet-standard sentinel context; Cerberus secret deployment folded into new-repo setup. Cerberus PR Governance
• Dependabot pipeline v1 — a test-then-approve skill that installs, runs the tests, and only then approves a dependency PR. Dependabot Pipeline
• Session lifecycle — plan-state handoff, post-plan-write hooks (ADR-0215), and skills tracked in-repo with onboard sync.

May — Orchestration + Governance at Fleet Scale

The peak month — more shipped here than in the entire January-to-March window.

• Orchestrator maturation — retry now iterates on the prior draft instead of restarting; repo-aware end-to-end runs; commit checkpoints between pipeline stages. End-to-End Orchestration
• Skills → live-reload architecture — commands migrated to hot-reloadable skills; /onboard reduced to a thin shim over a deterministic pickup decider; per-file pickup-read logging; same-window /respawn.
• Dependabot pipeline v2 — cross-repo parallelism, fleet mode, Windows Task Scheduler integration, single-command overnight status.
• Governance at fleet scale — fleet-wide branch-protection audit and remediation, auto-merge readiness checks, a two-tier mutation-flag safety rule, and canonical-label backfills across every repo.
• Per-repo CLAUDE.md discipline (ADR-0219) — a lean per-repo template plus a drift linter that keeps every repo's rules in sync.

June — Governance-Client Migration + Orchestrator Lifecycle

Maturation, plus a major transport migration.

• Antigravity (agy) CLI migration (ADR-0220) — the governance client moved entirely to an OAuth/subscription transport; the paid API-key path was removed outright.
• Orchestrator artifact lifecycle (ADR-0221) — a terminal cleanup stage that merges the design PR, deletes landed copies, and removes worktrees automatically; hallucinated-API-call detection added to spec validation.
• Dependabot fleet backfill — dependabot.yml backfilled across every existing repo and emitted for new repos by the scaffolder.
• CloudFlare Access-protected app reference architecture (Standard 0022).

---

What's New (March 2026)

206 issues closed and 182 commits since the last major wiki update. Seven capability themes shipped:

Theme	What Shipped	Evidence	Wiki Page
Security Hardening	Secret guard hooks (12/17 bypass vectors blocked), fleet-wide deployment, bash-gate for destructive git	#663–#714	Secret Guard Architecture · Fleet Security Audit
Governance Automation	Cerberus GitHub App auto-reviewer, pr-sentinel commit standards, branch protection across 62 repos	#736–#749	Cerberus PR Governance · Branch Protection at Scale
Prompt Economics	System prompt caching (90% savings on repeated context), Haiku routing, batch generation, per-call cost tracking	#625, #641–#647	Prompt Economics
Platform Engineering	Bypassed Windows' 32,767-char process limit via temp-dir CLAUDE.md hack	#787	The WinError 206 Hack
Workflow Reliability	Structured output migration (regex → JSON schema), worktree isolation enforcement, WorkspaceContext refactor	#774–#842	Worktree Isolation · Workflow Reliability
GTM Readiness	Zero-human demo pipeline: blueprint abstraction, presentation TUI, Playwright capture, deterministic seeding	#812–#820	Zero-Human Demo Pipeline
Cost Optimization	implement_code.py split (1,814→7 modules), tiered retry pruning, $5 budget guard, circuit breaker	#655, #640–#646	Prompt Economics

---

The Problem We Solve

AI coding assistants like Claude Code and GitHub Copilot are transforming software development. But enterprise adoption stalls because:

Challenge	Reality
No coordination	Multiple agents conflict and duplicate work
No governance	Security teams can't approve ungoverned AI
No verification	AI-generated code goes unreviewed
No metrics	Leadership can't prove ROI
No cost control	Token costs spiral without budgets or circuit breakers
Permission friction	Constant approval prompts destroy flow state

Organizations run pilots. Developers love the tools. Then adoption plateaus at 10-20% because the infrastructure layer is missing.

---

The Solution

Capability	What It Does	Enterprise Value
Multi-Agent Orchestration	12+ concurrent agents, one identity	Scale without chaos
End-to-End Pipeline	Triage → LLD → Spec → TDD → PR (automated)	Issue to PR in hours
Gemini Verification	AI reviews AI before humans approve	Quality gates that work
Governance Gates	Enforced checkpoints (design, code, docs)	Security team approval
Cost Management	Per-call tracking, budgets, circuit breakers, prompt caching	Predictable spend
Secret Guard Architecture	Hook-based secret protection, bypass analysis, fleet deployment	Defense in depth
Cerberus Auto-Reviewer	GitHub App for autonomous PR governance	No human bottleneck
Observability	Telemetry, audit trails, structured LLM logging	Full visibility
Safety & Guardrails	Kill switches, cascade prevention, rollback	Responsible deployment
Permission Management	Eliminate friction, track patterns	Developer productivity
34 Audits	Security, privacy, AI safety, compliance	Compliance readiness

---

Key Differentiators

1. Multi-Model Adversarial Verification

Claude builds code. Gemini reviews it. This isn't "two models" — it's adversarial verification where one AI checks another's work before humans approve. Learn more →

2. Secret Guard Architecture

Two-layer hook system protecting secrets from AI agent session transcripts. 12 of 17 identified bypass vectors blocked; 5 remain provably unblockable (Turing-complete). Learn more →

3. Prompt Economics at Scale

System prompt caching (90% savings on repeated context), Haiku routing for boilerplate, batch generation, $5 budget guards, circuit breakers. Learn more →

4. Autonomous PR Governance

Cerberus GitHub App auto-approves PRs after pr-sentinel validation passes. No human bottleneck, no self-approval workarounds. Fleet-deployed across all repos. Learn more →

5. Platform Engineering Under Hard Constraints

Windows' 32,767-character process creation limit blocks large LLM prompts. The fix: write the system prompt as CLAUDE.md in a temp directory with a fake .git marker. Prompt caching preserved, zero overhead. Learn more →

6. Self-Improving Governance

The system learns from Gemini verdicts to improve templates automatically. 164+ verdicts analyzed, 6 template sections added. Learn more →

7. Friction-First Approach

Permission friction is the #1 adoption killer. Dedicated friction logging protocol (Zugzwang) identifies patterns and auto-remediates. Learn more →

---

Core Workflows

AssemblyZero implements a five-stage governed pipeline:

End-to-End Pipeline

graph LR
    T["Triage"] --> L["LLD"]
    L --> S["Spec"]
    S --> I["TDD Impl"]
    I --> P["PR"]

Loading

Issue → Triage → Design → Spec → Code + Tests → Pull Request. Each stage is retryable, resumable, and gated. Learn more →

Requirements Workflow

graph TD
    I["Issue Created"]
    L["Write LLD"]
    G{"Gemini<br/>Review"}
    R["Revise"]
    A["APPROVED"]
    C["Ready for<br/>Implementation"]

    I --> L
    L --> G
    G -->|"BLOCK"| R
    R --> G
    G -->|"APPROVE"| A
    A --> C

Loading

Design documents are reviewed by Gemini before any code is written. Learn more →

Implementation Workflow

graph TD
    S["Start Coding"]
    W["Create Worktree"]
    I["Implement"]
    T["Run Tests"]
    R["Generate Reports"]
    G{"Gemini<br/>Review"}
    P["Create PR"]
    M["Merge & Cleanup"]

    S --> W
    W --> I
    I --> T
    T --> R
    R --> G
    G -->|"BLOCK"| I
    G -->|"APPROVE"| P
    P --> M

Loading

Code is reviewed by Gemini before PR creation. Worktrees enforce isolation — each issue gets its own sandbox. Learn more → · Worktree Isolation →

---

The Cast

Persona	Function	Philosophy	Status
The Great God Om	Human Orchestrator	Pure Intent	Active
Moist von Lipwig	Pipeline Orchestration	Keep messages moving	Implemented
Lord Vetinari	Work Visibility	Information is power	Planned
Brutha	RAG Vector Store	Perfect recall	⚠ Retired 2026-07-06 (ADR-0223)
The Librarian	Document Retrieval	Protect the books	⚠ Retired → Tiphys (ADR-0223)
Hex	Codebase Intelligence	Process. Compute. Return.	Vector RAG retired; AST grounding → Tiphys (ADR-0223)
The Historian	Duplicate Detection	History is a responsibility	Implemented
Captain Angua	External Intelligence	Sensory awareness	Implemented
Lu-Tze	Repository Hygiene	Constant sweeping	Implemented
DEATH	Doc Reconciliation	INEVITABLE. THOROUGH.	Manual
Commander Vimes	Regression Tests	Deep suspicion	Planned
Cerberus	PR Governance	Three heads, one gate	Implemented

Full cast →

---

Architecture

Document	Content
System Overview	Persona map, layer diagram, implementation status
Data Flow	Pipeline flow (Brief → PR), RAG indexing/retrieval
Workflow Interactions	Workflow chaining, checkpointing, human gates
ADR-0210: Persona Convention	Naming rules, when to create new personas
ADR-0211: RAG Architecture	Brutha foundation, Librarian/Hex/Historian consumers
ADR-0212: Local-Only Embeddings	Privacy rationale, model choice

---

Roadmap

Timeline	Milestone	Status
Q1 2026	LangGraph state machines, checkpointing	SHIPPED
Q1 2026	End-to-end orchestration pipeline	SHIPPED
Q1 2026	Cost management & circuit breakers	SHIPPED
Q1 2026	Prompt caching & model routing	SHIPPED
Q1 2026	Secret guard architecture	SHIPPED
Q1 2026	Cerberus PR governance (GitHub App)	SHIPPED
Q1 2026	Fleet-wide security audit & remediation (62 repos)	SHIPPED
Q1 2026	Structured output migration	SHIPPED
Q2 2026	Dependabot governed-dependency pipeline (fleet-wide)	SHIPPED
Q2 2026	Fleet governance at scale (branch protection, auto-reviewer)	SHIPPED
Q2 2026	Governance client → OAuth/subscription transport (ADR-0220)	SHIPPED
Q2 2026	Orchestrator artifact lifecycle & auto-cleanup (ADR-0221)	SHIPPED
Q2 2026	Zero-human demo pipeline (BoostGauge)	In Progress
Q2 2026	Supervisor pattern, LangSmith observability	Planned
Q3 2026	Dynamic tool graphs, multi-tenant support	Planned

See: LangGraph Evolution for the full technical vision.

---

Get Started

1. Find your path — Start Here (top of this page)
2. Read the architecture — Multi-Agent Orchestration
3. See the pipeline — End-to-End Orchestration
4. Understand cost control — Prompt Economics
5. See the velocity — March 2026 Velocity Report
6. Try it — Quick Start

---

"A man is not dead while his name is still spoken." GNU Terry Pratchett