@agenticprimitives/intent-marketplace

Part of Agentic Primitives — the open-source trust substrate for agentic applications: identity that can sign, authority checked at act time, evidence the owner carries. Developer kit · All packages

Status: STUB (Wave 0.5 of the W1 implementation wave). Typed primitives and matcher math ship today; the full Direct Lane marketplace lands in Wave 5 per the w1 implementation wave plan.

Agents are about to negotiate with each other at scale, and the intent layer is where that negotiation starts: "I want X, under these constraints — find me a counterparty." ERC-7683 and Anoma proved the pattern for cross-chain settlement; what no one ships is intents that live under the same delegation, custody, and audit substrate as the identity expressing them. An intent here is expressed by a canonical Smart Agent address, brokered under a scoped delegation, and resolved with a signed, AI-auditable receipt — not a freeform message in a vendor silo.

This package is the designed coordination slice of that substrate: spine Layers 2, 3, 5, 6, and 7 — intent expression through dual-signed commitment — spec'd in full, scaffolded now, landing in the implementation waves.

Part of agenticprimitives — the trust substrate for the agent economy: one canonical Smart Agent identity with custody, delegation, naming, credentials, and audit evidence designed as one system.

What ships today

The W1 foundational surface — types and pure functions, no marketplace runtime:

• Intent typed envelope — direction (receive | give) + SKOS object IRI + topic + payload + expected outcome + 5-tier visibility + status state machine.
• ConstraintSet + AssumptionSet — first-class typed structures (Anoma CSP-shaped), never freeform payload. Every constraint carries its source (user-asserted | llm-inferred | policy-imposed), strength, and enforcement point.
• ResolutionReceipt type — the AI-auditability surface: model + version + prompt hash + tool-call hashes + confidence + a requiresUserConfirmation gate, so an LLM-inferred constraint can never silently bind a user.
• IntentMatch + Commitment types — the match record and the dual-signed envelope handed off (as bytes) to agreements.
• Matcher math — isCompatible (opposite direction + same object + topic similarity) and the composite score (0.6 × proximity + 0.4 × outcome, Laplace-smoothed) with toMatchScore.

What lands in the waves (designed, not shipped)

buildIntent / signIntent, the runMatcher / rankCandidates pipeline, visibility projections (projectFor with per-field DisclosurePolicy), the three-tier delegation model (T1 session / T2 system / T3 cross), and the scope catalog + cross-delegation builders. Pool Lane (1-to-N) and Proposal Lane (RFP) are deliberately deferred beyond W1. The resolver engine is a separate package — intent-resolver.

The engagement layer (spec 336) is deliberately NOT this package. The pre-task provider-qualification chain — audience-bound IntentProjections, EngagementProbe + the six-kind response taxonomy, provider-signed FulfillmentOffers, RAR-shaped MandateRequirements, the deterministic engagement reducer — is the dedicated intent-engagement package (spec 336 §8; wave E1). It sits above this one and imports our intent types type-only; this package matches counter-intents within one trust context, engagement qualifies external cross-org providers. See spec 336 + the standards crosswalk.

Where this is heading / market context

• ERC-7683 standardized cross-chain intent orders and the resolver role; we adopt its resolver/matchmaker separation and reserve an erc7683Order slot for interop.
• Anoma shaped intents as constraint-satisfaction problems; our ConstraintSet follows that shape as a first-class type.
• x402 and Google AP2 are making machine payments and agent mandates real — which is exactly why the intent that triggers a payment must share a trust chain with the mandate that authorizes it. Here, the intent, the delegation that brokered it, the payment mandate it binds to (payments), and the audit event it emits all reference one canonical address.

That combination — intents under the same delegation + custody + audit substrate — is the gap this package exists to fill.

Authoritative spec: spec 239 — intent spine (see spec.md). Owns spine layers 2-3-5-6-7; bounded surface in CLAUDE.md and capability.manifest.json.

Build

pnpm --filter @agenticprimitives/intent-marketplace typecheck
pnpm --filter @agenticprimitives/intent-marketplace test
pnpm --filter @agenticprimitives/intent-marketplace build

Status — honest version

STUB. Treat every API outside "What ships today" as a design commitment, not a capability. Wave sequencing: w1-implementation-wave-plan.md. Security findings for the whole repo are tracked live in docs/audits/findings.yaml.