A2AC

Verifiable Coordination For AI Agents

Tasks, receipts, approvals, and artifact integrity for multi-agent work.

Website · Cube Protocol · Phil Hills

---

What A2AC Is

A2AC is an operating layer for AI agents that need to coordinate real work.

It gives agents a shared way to:

• receive tasks
• route work to the right model
• produce durable results
• sign execution receipts
• preserve context across sessions
• request human approval for high-risk actions
• verify artifacts after the chat or process is gone

The core idea is simple: agent work should leave evidence.

Why It Matters

Most agent systems still treat chat output as the source of truth. That breaks down quickly when multiple models, tools, files, approvals, and background jobs are involved.

A2AC treats agent execution as operational state:

• Tasks define what should happen.
• Receipts prove what happened.
• Results preserve durable output.
• Approvals capture human authority.
• Artifacts can be hashed, signed, and verified.

This makes agent work auditable, resumable, and easier to coordinate across models.

Core Concepts

Task Files

Agents work from structured task specifications instead of loose chat memory.

.a2ac/tasks/
  task-id.json
  task-id-result.md

Signed Receipts

Meaningful actions produce receipt records that can be verified later.

{
  "receipt_id": "e577c283e63f",
  "type": "execution_receipt",
  "timestamp": "2026-07-02T22:41:07Z",
  "agent": "codex",
  "task": "Build receipt verifier",
  "signature": "43d3777067cce208...",
  "prev_hash": "ef64743881165dcb..."
}

Human Approval Gates

High-risk actions should be explicit approval objects, not implied consent hidden in chat.

Examples:

• production deploys
• billing changes
• public publishing
• security boundary changes
• destructive file or cloud operations

Multi-Model Coordination

A2AC is model-neutral. Different agents can specialize:

• planning and orchestration
• code implementation
• browser/CDP tasks
• review and verification
• local deterministic automation

The goal is not one model doing everything. The goal is shared state, clear routing, and verifiable handoff.

Artifact Integrity

A2AC uses Cube Protocol concepts for portable artifact integrity:

• RFC 8785 JSON canonicalization
• content hashes
• signatures
• schema references
• provenance links

See cube-protocol-spec for the public reference profile.

Example Workspace Shape

.a2ac/
  tasks/            task specifications and result files
  approvals/        pending, approved, and rejected human decisions
  notifications/    agent and task events
  projects/         grouped workstreams
  task_agent/       active focus and routing briefs

evidence/
  receipts/         signed receipt records

Design Principles

• Disk is ground truth. If it is not written and verifiable, it did not happen.
• Agents should not rely on memory. Work must survive restarts, crashes, and lost chats.
• External AI can advise. Trusted local agents perform state-changing execution.
• Approvals are first-class. Human judgment should be captured as signed state.
• Verification beats narration. Claims should be checked against files, receipts, and hashes.

Status

A2AC is an active implementation and protocol workspace.

The public repositories are being cleaned and separated into:

• core coordination concepts
• artifact integrity profile
• extension and UI surfaces
• examples and SDKs

Related Work

• Cube Protocol Spec
• A2A Content Integrity Proposal #1140
• startq

License

Apache-2.0. See LICENSE.

---

Built by Phil Hills, founder of A2AC LLC.