DAREP

Working Note · Milestone A0 · November 2026 · Open for Critique

Why agents need durable, verifiable work delivery—and where a small open protocol belongs

Before freezing our wire v2 specification, we surveyed the closest agent protocols and federated systems. This note records what we adopt, what we bridge, what nobody specifies or ships yet, and what would change our mind.

Position. Keep a small protocol of our own—three short documents covering signed records, a delivery contract, and one HTTPS binding—so the carrier can change without invalidating stored evidence. Meet MCP at the agent edge, and bridge A2A and email at an organization’s public door.

1. Where each boundary sits

Boundary Mechanism Role
Agent to provider MCP (pull-only) Small fixed tool set with request keys for idempotent starts and scoped, revocable grants. No reliance on server push.
Provider to provider DAREP Compact JWS records over HTTPS POST. Verify, persist durably, then return a signed storage receipt.
Public doorways A2A & Email gateways Planned optional edge extensions. Work identifiers and states are designed to map cleanly onto A2A without altering core records.
Address & keys acct: today; WebFinger + W3C CID planned Today kid is the acct: address. Wire v2 plans to carry the address and a key URL (kid) so resolution rules can evolve independently of stored signatures.
Evidence & settlement Self-contained JWS + reserved hook Receipts and work events verify offline today. Grants and revocations are local provider records until a signed delegation record is specified. A reserved hook is planned so x402 or escrow can attach later with zero wallet in core.

2. What nobody specifies or ships yet

Single-team agent messaging and “an email inbox for an agent” are already crowded. Meanwhile, signed federated envelopes have been specified on paper several times without gaining independent implementers. What remains genuinely missing across the field are four concrete primitives:

  • Separate signed work facts: distinguishing delivered (the storage receipt), accepted (against a specific terms version), completed, and verified as distinct cryptographic events.
  • Third-party verifiable evidence: A2A signs only Agent Cards; ANP signs only group results; AMP’s receipts are unsigned. None preserves an offline-verifiable work chain.
  • Owner-approved context by reference: versioned, access-controlled artifacts without copying private memory across organizational boundaries.
  • Optional settlement attached to verifiable work: payment and trust drafts (x402, ERC-8004, ERC-8183) all assume a durable messaging layer already exists beneath them.

3. Why not build inside an existing protocol?

Candidate Observed state Decision
A2A v1.0.1; Apache-2.0; broad foundation support. Assumes a reachable server, makes idempotency optional, has no message sender field, and states messages must not be treated as reliable delivery for critical information. Bridge at an organization’s public edge. Keep UUIDs and task vocabulary mechanically mappable.
MCP Revision 2026-07. Stateless tool access with no durable redelivery to disconnected runtimes. Keep at the agent-to-provider edge, pull-only.
ANP v1.2. Direct-messaging profile states success does not guarantee persistence and forbids requiring retries; single active committer in recent history. Borrow identity ideas. Do not adopt as base.
AMP Draft v0.1.2, no releases, unsigned receipts, lacks multi-party threads and work events. Borrow ideas. Do not adopt as base.
Matrix Mature federated room-state replication; heavier shared-state machinery than asynchronous work handoffs need. Architectural reference only.
Email (SMTP) Proven store-before-ack federation, but identity is tied to the host and it lacks native idempotency keys, typed work events, and signed receipts. Copy the delivery contract; add an email gateway as an edge extension.

4. Consequences for wire v2

Wire v2 is designed to be our single breaking change. It freezes only the structural envelope and receipt contract; roles, work lifecycles, admission controls, and gateways ship as versioned URI extensions under https://darep.org/ext/....

  • Key URLs & rotation: kid becomes a key URL referencing an identity document with purpose-separated P-256 keys, validity windows, and retained history so old signatures stay verifiable.
  • Immutable bytes on retry: outgoing signed JWS records are stored once and resent byte-for-byte; inbound signed envelopes and issued receipts are retained permanently as evidence.
  • Verify, store, acknowledge, then apply: a provider can durably store and issue a signed receipt before running any extension workflow logic—and with zero model calls.
  • Closed, never deleted: owner approvals and agent grants are closed with timestamps so post-hoc audits always know who acted and under what authority.

5. Core, extensions, and commercial boundaries

Commercial products, including hosted providers, get no private path. Any capability a commercial product needs enters the public specification first.

Never charged for as a condition of the protocol: running a provider, interoperating across domains, basic authorization, exporting history, migrating identities, ordinary public indexing, and choosing third-party verification or settlement services.

6. What would change this position

  • If A2A adopts mandatory idempotency, message signatures, and durable store-and-forward semantics, we would promote our A2A binding to the primary carrier.
  • If two prospective provider operators tell us our native HTTPS wire is what prevents adoption—or if a major vendor ships cross-owner durable messaging—we will narrow DAREP to the work-evidence and receipt layer over that carrier.
  • Conversely, one independent implementation passing our upcoming black-box conformance kit validates the minimal core.