APAAI Protocol v1.0 (RFC-2025-001)

APAAI — the Accountability Protocol for Agentic AI — defines a standard way to record autonomous actions as Action → Policy → Evidence.

Open standard for accountable autonomous systems.

APAAI provides a model-agnostic accountability layer for agentic AI, defining how intent, policy, and evidence are expressed and verified — enabling organizational governance without prescribing workflows.

Apache-2.0 License • Maintained by apaAI Labs

Protocol Specification

APAAI v1.0 specifies a vendor-neutral HTTP API and data model for auditable autonomous actions. It standardizes how agents propose actions, evaluate policies (enforced or observed), and emit verifiable evidence records — providing the substrate that organizational governance can build upon.

Model- and vendor-agnostic HTTP API
Portable policy definitions (enforce or observe)
Human-in-the-loop approval as a first-class primitive
Verifiable evidence records (signing optional)

Protocol Flow (Python)

# Action → Policy → Evidence

from apaai import AccountabilityLayer

# initialize accountability layer
apaai = AccountabilityLayer(endpoint="http://localhost:8787")

# 1) Propose an action (intent)
decision = apaai.propose(
    type="send_email",
    actor={"kind": "agent", "name": "mail-bot"},
    target="mailto:client@acme.com",
    params={"subject": "Proposal"}
)

# 2) Evaluate/enforce policy (HiTL if required)
if decision["status"] == "requires_approval":
    apaai.human.approve(decision["actionId"], approver="@reviewer")

# 3) Add verifiable evidence
apaai.evidence.add(decision["actionId"], [
    {"name": "email_sent", "pass": True, "note": "msgId=123"}
])

Data Models

Three primitives define the accountability loop.

Action

Structured intent (who/what/target/params/timestamp).

{
  "id": "a_123",
  "type": "send_email",
  "actor": { "kind": "agent", "name": "mail-bot" },
  "target": "mailto:client@acme.com",
  "params": { "subject": "Hi" },
  "timestamp": "2025-10-10T12:34:56Z"
}

Policy

Rules that constrain execution; may enforce or observe.

rules:
  - when: { actionType: "send_email" }
    require: ["reviewer_approval"]
    mode: enforce

Evidence

Attestable outcomes others can verify independently.

{
  "actionId": "a_123",
  "checks": [{ "name": "reviewer_approval", "pass": true, "approver": "@fernando" }],
  "timestamp": "2025-10-10T12:40:00Z"
}

API Reference

Machine-readable OpenAPI schema and a human-readable reference.

OpenAPI Spec Download YAML GitHub

Governance

APAAI defines accountability primitives; governance is intentionally out of scope. The protocol enables governance by ensuring consistent per-action records of intent, applied policy, and evidence. apaAI Labs stewards the reference implementations and RFC process.