agent-service-agreements

Agent Service Agreements

Machine-readable service contracts and quality verification for autonomous agent commerce. Part of the Agent Trust Stack.

from agent_service_agreements import Identity, create_agreement_from_template, VerificationEngine

agreement = create_agreement_from_template("research",
    client=Identity(scheme="api_key", value="client-123"),
    provider=Identity(scheme="api_key", value="provider-456"),
    escrow_amount="5.00")

result = VerificationEngine().verify(deliverable_text, original_request="Summarize federated learning research")
print(f"Score: {result.composite_score}, Passed: {result.passed}")

What It Does

ASA provides two API surfaces for agent-to-agent commerce:

Agreements API — Create, negotiate, sign, and manage machine-readable service contracts with quality criteria, SLO definitions, and payment terms.

Verification API — Evaluate deliverable quality across multiple dimensions (correctness, completeness, coherence, etc.) with structural and semantic verification. Works standalone or against agreement criteria.

Install

pip install agent-service-agreements

Zero required dependencies. Python 3.8+.

Optional integrations:

pip install agent-service-agreements[arp]  # Agent Rating Protocol
pip install agent-service-agreements[ajp]  # Agent Justice Protocol

CLI

# Create an agreement from a template
agent-service agree --template research --client alice --provider bob --amount 5.00

# Verify a deliverable
agent-service verify --deliverable output.md --request "Summarize FL research" --type text/research

# List templates
agent-service templates

# Check store status
agent-service status

Agreement Templates

Escrow & Graduated Payment

ASA supports graduated payment release based on quality scores:

Continuous mode available: payment = (score / 100) * amount.

Dead-man's switch defaults to hold_for_backup_evaluator (not release_to_provider).

Negotiation

Multi-round structured negotiation with fairness constraints:

• Configurable round limits (default: 5)
• Asymmetry limits (max 25% change per round)
• Price bounds relative to market rates
• Structured JSON messages (no free-text manipulation)

Verification Depths

• Structural: Schema validation, format checks, size constraints (~0ms)
• Semantic: Agent-as-a-Judge integration point for content quality (~10-120s)
• Composite: Structural + semantic + optional canary/cross-reference checks

Trust Ecosystem

ASA is Layer 2 in the Agent Trust Stack:

Layer 3: Accountability  — Agent Justice Protocol (disputes, forensics)
Layer 2: Agreements      — Agent Service Agreements (this package)
Layer 1: Trust Primitives — Chain of Consciousness + Agent Rating Protocol

Security

VAM-SEC v1.0 Security Disclaimer: This is alpha software implementing a protocol specification. It has not undergone formal security audit. Do not use for production financial transactions without additional security review. The escrow abstraction layer does not handle actual funds — it computes release amounts that must be executed by an external payment system.

License

Apache 2.0. See LICENSE.

Links

• Whitepaper
• Agent Rating Protocol
• Agent Justice Protocol