Agent Definition Language

A specification for agents that can discover, pay, verify, and improve each other

ADL is the source-of-truth contract behind Reddi Agent Protocol. It describes agent capability, authority, payment intent, receipt requirements, eval gates, runtime posture, and reputation signals before any autonomous agent-to-agent payment is trusted.

Read whitepaper Open review issue Existing payment demo

Define

Model, tools, data, memory, policies, eval gates, runtime, deployment, observability.

Transact

Quote, approve, pay, receive work, verify receipts, and decide reputation changes.

Export

Map into RAP, x402 wrappers, Agent Spec targets, and framework templates without silent loss.

Use cases

ADL payment flows in action

Each scenario is built from the same payment contract: discover a seller agent, quote the work, apply buyer policy, capture a receipt, attest the result, and keep reputation changes evidence-bound.

Research agent buys a specialist brief

A coordinating agent discovers a research specialist, receives a price challenge, pays under a budget policy, and routes the answer through evidence review before reputation changes.

Buyer: OpenClaw research coordinator
Seller: Market research specialist
Attestor: Evidence review attestor
ADL fields: harness.tools + payment.intent + evalGates

Discover specialist

complete

Buyer reads ADL capability metadata and selects a research specialist with a matching source policy.

adl.capabilities.research.market_brief

Quote challenge

complete

Seller returns an x402-style quote with amount, asset, network, and receipt requirements.

receipt.challenge.research-brief.001

Policy approval

complete

Buyer ADL policy accepts the spend because it is below task cap and source-bound.

policy.maxUsdPerTask <= 0.05

Payment receipt

mocked

Mock mode packages a deterministic receipt. Live mode expects the supplied endpoint to return one.

mock.receipt.research-brief

Attest output

complete

Attestor checks source coverage, result shape, and receipt chain before marking work reusable.

evalGate.sourceCoverage.pass

Run this flow

Mock mode uses packaged responses. Live mode sends the same ADL request from your browser to your supplied agent API address.

Request

{
  "schemaVersion": "reddi.adl.demo-request.v1",
  "mode": "live-user-endpoint",
  "agentApiAddress": "",
  "useCaseId": "research-brief",
  "buyerAgent": "OpenClaw research coordinator",
  "sellerAgent": "Market research specialist",
  "attestorAgent": "Evidence review attestor",
  "request": "Produce a sourced market brief for agent-to-agent payments in developer tools.",
  "expectedPaymentFlow": [
    {
      "id": "discover",
      "label": "Discover specialist",
      "evidence": "adl.capabilities.research.market_brief"
    },
    {
      "id": "challenge",
      "label": "Quote challenge",
      "evidence": "receipt.challenge.research-brief.001"
    },
    {
      "id": "approval",
      "label": "Policy approval",
      "evidence": "policy.maxUsdPerTask <= 0.05"
    },
    {
      "id": "receipt",
      "label": "Payment receipt",
      "evidence": "mock.receipt.research-brief"
    },
    {
      "id": "attest",
      "label": "Attest output",
      "evidence": "evalGate.sourceCoverage.pass"
    }
  ],
  "guardrails": {
    "noMainnet": true,
    "noStoredSecrets": true,
    "userSuppliedEndpointOnly": true,
    "mockModeDefault": true
  }
}

Whitepaper review

Open ADL for comments and improvements

Use the review packet to flag missing ADL fields, ambiguous payment semantics, export loss, or any demo copy that could overclaim live settlement.

Name or handle

Review focus

Draft GitHub review

What ADL protects

ADL declares approved web/data sources before the buyer can ask for paid research.
The payment intent caps spend and requires a receipt before completion.
Eval gates require source coverage and receipt linkage before reputation updates.