XAP defines how autonomous agents discover counterparties, negotiate contracts, execute settlements, and produce replayable receipts that explain what happened. Humans review outcomes after the fact.
AI agents can plan, code, and orchestrate complex workflows. But when agents need to do business with each other, there is no shared language for how that happens.
Existing protocols assume a human approving the transaction. XAP is designed for a world where agents negotiate with agents, and humans review the outcome later.
| Protocol | Model |
|---|---|
| Stripe/OpenAI ACP | Agent-assisted shopping |
| Google AP2 | Human-authorized payments |
| Coinbase x402 | Pay-per-request API access |
| XAP | Autonomous agent-to-agent |
Every platform defining its own negotiation, settlement, and verification flow creates fragmentation. A protocol creates a shared language so that any agent system can transact with any other system.
XAP defines that language.
Verity is the deterministic truth engine that records decisions, policy snapshots, evidence references, and replayable outcomes.
Given the same inputs, replay always produces the same result. This is how agents prove what happened and why.
agnt_Cryptographic identity for an autonomous agent. Contains public keys, declared capabilities, pricing model, and historical performance metrics.
neg_Structured offer/counter-offer/accept flow between agents. Defines terms, pricing, SLAs, and expiration. Every state transition is signed.
stl_Declaration that funds are locked against conditional release. Supports multi-party splits, payment hold conditions, and failure modes.
rcp_Immutable proof that a settlement executed on a rail. Contains adapter metadata, transaction references, and finality timestamps.
vry_Deterministic explanation of why a decision resolved the way it did. Policy snapshot, evidence references, computed outcome. Fully replayable.
Agent registers identity and declares capabilities
Agent discovers counterparties on the network
Agents exchange OFFER / COUNTER / ACCEPT
SettlementIntent created, funds held pending verification
Settlement executes on rail adapter
ExecutionReceipt issued with tx reference
VerityReceipt recorded with policy snapshot
Outcome is deterministically reproducible
Structured OFFER → COUNTER → ACCEPT flow. Agents negotiate terms in real time without human intervention.
The primitive is not 'pay'. It's 'lock funds, evaluate condition, release or refund'.
Distribute value across multiple agents atomically. Defined in basis points, executed as one transaction.
Every settlement captures why it resolved the way it did. Given the same inputs, replay produces the same result.
SUCCESS, FAIL, UNKNOWN, DISPUTED, REVERSED. UNKNOWN is a first-class state. No false confidence.
Agents accumulate permanent execution history. Trust becomes computable rather than assumed.
OFFER → COUNTER → ACCEPT
OFFER → REJECT
COUNTER → ACCEPT
COUNTER → REJECT
* → EXPIRED
PENDING → EXECUTING → SUCCESS
PENDING → EXECUTING → FAIL
PENDING → TIMEOUT → REFUND
* → DISPUTED
XAP defines the schema and behavior. Verity provides deterministic truth. The Settlement Engine executes conditionally. Adapters connect to any rail.
Each layer is independently implementable. Build your own stack, or use the reference implementation.
ZexRail is the reference implementation of XAP — the fastest way to get to production.
Settlement Rails
{
"negotiation_id": "neg_01JD8K2M4N5P6Q7R8S9T0U1V2W",
"initiator": "agnt_payment_processor",
"responder": "agnt_inventory_manager",
"state": "OFFER",
"terms": {
"action": "reserve_inventory",
"quantity": 100,
"unit_price": { "amount_minor": 2500, "currency": "USD" }
},
"conditional_hold_required": true,
"expires_at": "2025-01-15T15:00:00Z"
}{
"settlement_id": "stl_01JD8K2M4N5P6Q7R8S9T0U1V2W",
"negotiation_id": "neg_01JD8K2M4N5P6Q7R8S9T0U1V2W",
"amount": { "amount_minor": 250000, "currency": "USD" },
"splits": [
{ "agent_id": "agnt_vendor", "bps": 7000 },
{ "agent_id": "agnt_platform", "bps": 2500 },
{ "agent_id": "agnt_referrer", "bps": 500 }
],
"conditional_hold": {
"release_on": "delivery_confirmed",
"refund_on": "delivery_failed"
}
}Every state transition emits an event. Subscribe to the events you care about. Build reactive agent systems.
agent.registered
negotiation.created
negotiation.accepted
settlement.pending
settlement.executed
receipt.generated
replay.completed
dispute.opened
Objects are signed by the issuing agent. Signatures are verifiable.
State machines enforce valid transitions. Invalid states are rejected.
Every operation has a unique key. Duplicates are safely ignored.
Execution outcomes are recorded permanently. History cannot be rewritten.
Any decision can be replayed. Same inputs produce same outputs.
Agent reputation accumulates over time. Trust becomes computable.
Implemented by
Read the specification, explore the reference implementation, or join the community on GitHub.