ReddiAgent ADL

The open Agent Definition Language — define an AI agent once: model envelope, operating harness, and payment authority. Validate it deterministically. Run it anywhere.

v0.2.0-betaApache-2.0 codeCC BY 4.0 specs235-test suite5-level conformance

The thesis

Agent = model definition + harness definition + settlement/reputation extension

Useful agents shouldn't start as vendor-specific scripts. ADL describes the job, model needs, tools, policies, eval gates, observability, and payment intent once — then compatibility reports tell you exactly what survives on each target (Microsoft Agent Framework, Google ADK, LangGraph, OpenAI, Anthropic, Gemini, Ollama…): supported, degraded, or unsupported. Nothing silently dropped.

Settlement is not success

USENIX Security 2026 research found security-rule violations in all 15 major x402 payment facilitators. Our answer ships in this release: a receipt-integrity validator that refuses to treat payment proof as service success — receipts must bind delegated authority, payment evidence, service outcome, and eval evidence across 10 layers, tested against a 14-case threat model (replay, wrong-payee, paid-but-denied). Hardened by three rounds of adversarial review: 33 attack probes, every one failing closed.

The five protocol layers a receipt binds, and the role of each
x402payment challenge/proof — the rail
AP2 / Verifiable Intentdelegated authority — who allowed what
MCPprotected-resource access — what was touched
Solanasettlement proof — where value moved
RAPreceipts, accounting, reputation — bound above the rail

Agents paying agents — both sides, today

Buyer side

A paid-research agent with a full payment authority contract: principal, spender, cap, expiry, revocation, receipts — human approval before any spend.

Seller side

Three delegation services (research, pricing, review) other agents can discover and pay — charge intents, receipt binding, reputation signals.

Dry-run rail only

Every example runs on x402-dry-run. Live settlement stays gated behind external audit — fail-closed by design.

Download

The bundle: canonical spec + machine-checked schema + 16 validated examples + v0.1→v0.2 migration guide + the receipt-integrity validator and conformance checker (runnable: python3 scripts/adl_v02_conformance.py examples/v0.2/simple-agent.yaml).

Spec home

The specification lives at github.com/nissan/reddiagent-lab — issues and structured review intake included.

The spec repo is going public with tagged release v0.2.0-beta — if the GitHub link 404s, the flip is imminent. Everything on this page is downloadable right here in the meantime.
ReddiAgent Lab · code Apache-2.0 · specifications & docs CC BY 4.0 · Live payment rails and mainnet remain gated behind external security audit. Built in the open, reviewed adversarially, by agents — with a human holding the keys.