AgentCourt vs Arbitova: Choosing a Dispute Resolution Layer

June 23, 2026 · View on GitHub

Both AgentCourt and Arbitova solve the same problem: what happens when agent commerce transactions go wrong. They take fundamentally different approaches.

TL;DR

AgentCourtArbitova
ApproachPolicy-driven (rules engine)LLM-based (N=3 majority vote)
Speed<500ms~30s
DeterminismSame input = same rulingNon-deterministic
Cost$0.05 flat per dispute2% of transaction value
EscrowNot requiredRequired (smart contract)
AppealsBuilt-in appeal + human fallbackescalate_to_human flag
TransparencyFull rule-based reasoningLLM reasoning (opaque)
Blockchainx402 payments, verdict attestationFull on-chain escrow lifecycle
LicenseMITProprietary

When to Choose AgentCourt

  • You need fast, predictable rulings (API quality checks, SLA monitoring)
  • You want low cost dispute resolution ($0.05 vs 2%)
  • You don't want escrow (direct payments, reputation-based enforcement)
  • You need auditable, deterministic outcomes
  • You have structured evidence (booleans, numbers, dates)

When to Choose Arbitova

  • You need full escrow lifecycle (lock funds, deliver, confirm)
  • Disputes are nuanced/qualitative (content quality, creative work)
  • You want on-chain enforcement (smart contract custody)
  • You're building a marketplace with buyer/seller protection

When to Use Both

AgentCourt and Arbitova are complementary:

  1. AgentCourt handles clear-cut violations (schema mismatch, non-delivery, SLA breach)
  2. Arbitova handles nuanced disputes where judgment is needed
  3. Escalation chain: policy check → LLM arbitration → human review
Dispute filed
  → AgentCourt: policy-driven ruling (<500ms, \$0.05)
  → If confidence < 0.70 or ambiguous: escalate to Arbitova (LLM, 30s, 2%)
  → If still unclear: human review

Architecture Comparison

AgentCourt (Non-Custodial)

Agent A pays Agent B via x402 (direct, no escrow)
  → Transaction goes wrong
  → AgentCourt evaluates structured evidence
  → Returns deterministic ruling
  → Marketplace adjusts reputation / triggers refund

Arbitova (Escrow-Based)

Agent A locks USDC in Arbitova smart contract
  → Agent B delivers work
  → Agent A confirms or disputes
  → If disputed: Arbitova arbiter evaluates (LLM, 30s)
  → Smart contract releases/refunds based on verdict

Policy Templates

AgentCourt ships with 7 specialized policy templates:

  1. api-quality (7 rules): Schema mismatch, wrong format, stale data
  2. freelance-delivery (5 rules): Non-delivery, late delivery, partial
  3. milestone-payment (6 rules): Unpaid milestones, partial payments
  4. bug-bounty (5 rules): Reproducibility, severity disputes
  5. sla-monitoring (6 rules): Uptime, latency, availability
  6. scope-dispute (4 rules): Budget exceedance, unauthorized changes
  7. physical-commerce (6 rules): Damaged goods, wrong items, returns

Arbitova uses a single general-purpose arbitration model.

AgentCourt vs aubinhaba/dispute-resolution-agent

DimensionAgentCourtdispute-resolution-agent
ApproachDeterministic rules (no LLM for ruling)Multi-agent LLM (Spring AI)
LanguagePython (FastAPI)Java (Spring Boot)
DeterminismGuaranteed — same input → same outputProbabilistic — validation layer attempts to enforce consistency
Speed<500msMultiple seconds (LLM inference)
Cost$0.05/disputeLLM API cost per dispute
FocusAgent commerce (x402, API quality, SLA)Payment chargebacks (traditional finance)
RulesJSON policy templates, community-contributableRAG over rule corpus
Audit trailMatched rule ID + confidencecitedRulePassages + evidenceRefs
MCPNative MCP server (6 tools)MCP client for transaction data
APIREST + SDK + PostmanNot exposed as API
LicenseMITUnknown

Key insight: This competitor validates the market need for dispute resolution in automated transactions. Their approach (LLM + validation) is the opposite of ours (deterministic rules). Both agree on the importance of auditability, but we achieve it through simpler means — if the rule is deterministic, the audit is automatic.

AgentCourt advantage: No LLM means no hallucination risk in rulings, no per-dispute inference cost, and consistent reproducibility for trust scoring.

Market Ecosystem (June 2026)

AgentCourt doesn't exist in isolation. Here's how we fit into the emerging agent commerce stack:

LayerProjectWhat They DoRelationship
IdentityOpen Agent Trust RegistryRoot-of-trust for agent identity (Ed25519 attestations)Complementary — Their identity + our disputes = complete trust
GovernanceMicrosoft Agent Governance ToolkitPolicy enforcement, zero-trust for AI agents (4.4K★)Adjacent — Governance before transaction, disputes after
Paymentsx402 ProtocolHTTP 402 → USDC micropaymentsIntegration — We use x402 for per-dispute pricing
Audio LicensingResonateMachine-first audio licensing for agentic commercePotential customer — License disputes could use our API
LLM DisputesAI Dispute ResolutionEnterprise dispute resolution with AIDifferentiator — LLM-based vs our deterministic approach

The Trust Stack

┌─────────────────────────────────────────┐
│  Governance (MS Toolkit, etc.)          │ ← Pre-transaction rules
├─────────────────────────────────────────┤
│  Identity (Trust Registry, A2A)         │ ← Who are you?
├─────────────────────────────────────────┤
│  Discovery (Directories, Marketplaces)  │ ← How do I find you?
├─────────────────────────────────────────┤
│  Payment (x402, USDC)                   │ ← How do I pay you?
├─────────────────────────────────────────┤
│  ★ DISPUTE (AgentCourt) ★              │ ← What happens when it goes wrong?
├─────────────────────────────────────────┤
│  Reputation (Phase 2)                   │ ← What's your track record?
└─────────────────────────────────────────┘

We're the layer nobody else is building — and the one the stack needs most.