AgentCourt is a policy-driven dispute resolution API for agent commerce. When two AI agents transact and something goes wrong — late delivery, partial work, SLA breach, contested bug bounty — AgentCourt evaluates the evidence and produces a deterministic ruling.
Traditional dispute resolution is designed for human-initiated transactions with credit cards. Agent transactions happen at machine speed, involve non-traditional deliverables (code, data, API access), and need instant resolution. AgentCourt is purpose-built for this.
No. AgentCourt adjudicates based on policy rules and evidence, not legal frameworks. The ruling is a structured output (remedy + reasoning) that platforms can enforce programmatically.
Yes. The same evidence, submitted against the same policy template, will always produce the same ruling. We use pattern-based fact extraction and boolean rule evaluation — no LLM in the critical path.
Three reasons:
- Determinism — LLMs can produce different outputs for identical inputs
- Auditability — Our engine shows exactly which evidence contributed which weight
- Cost/latency — Our engine runs in <500ms at $0 marginal cost
Contracts, payment proofs, invoices, receipts, git commits, server logs, files, messages, screenshots, testimonials, and claims. Each type has a default weight (contract=1.0, screenshot=0.5, claim=0.1).
When evidence from both parties conflicts, the fact is resolved in favor of the higher-weighted evidence. The dispute is transparent in the ruling output — you can see both claims and why one prevailed.
No. AgentCourt is a ruling engine, not a financial intermediary. Platforms enforce rulings through their own escrow, reputation, or access control mechanisms.
Yes. Policy templates are JSON files. Define evidence weights, fact extraction patterns, and boolean rules. See CONTRIBUTING.md for details.
Average ruling generation: <500ms. The engine is stateless with no database dependencies.
Three options:
- REST API — POST to /v1/disputes with evidence, get ruling
- Python/JS SDK — Zero-dependency client libraries
- MCP Server — For agent frameworks that support Model Context Protocol
Yes. AgentCourt is protocol-agnostic. It evaluates disputes regardless of how the payment was made. The ruling output includes a structured remedy that any payment protocol can execute.
Yes. AgentCourt is MIT licensed. Clone the repo, run python -m uvicorn src.main:app, and you have your own dispute resolution engine.
The API is currently free during the design partner program (up to 100 disputes/day). Self-hosting is always free.
No. The engine is stateless. Dispute data is processed in-memory and returned in the response. We don't store evidence, contracts, or rulings (the calling platform tracks case history).
Yes. Evidence can include a content_hash field. When present, the engine adds a +0.10 weight bonus, recognizing that hashable evidence is more trustworthy.
All policy templates are open source and inspectable. Anyone can audit the rules before submitting a dispute. Custom templates let platforms define their own fairness criteria.