Agent interop protocols

The strongest input here is traffic data from a year of running a public tool server, and it is unflattering: handshakes dominate the wire, tool calls are a rounding error beside them, and several round trips happen before any service is rendered. Server-initiated back-channels need a permanently open connection that is very hard to hold on a public server, and sticky sessions make scaling painful. So the protocols are converging on stateless requests, cacheable declared capabilities discoverable without a handshake, and routing metadata infrastructure can read from headers without parsing bodies.

One camp bets that a deliberately tiny, schema-light envelope beats large-schema protocols: each endpoint can translate with its own intelligence and resolve ambiguity by asking rather than failing. The camp states its own boundary honestly — strict-schema protocols remain the right choice wherever governance, audit and enforcement matter. That is a choose-by-workload answer rather than a winner, and it is more useful than either partisan version.

Multi-agent systems running for hours to weeks need a provenance-carrying, governed representation of which context to hand over for which task — a gap neither loose markdown nor current agent-to-agent protocols close. Above that sits the open-web problem: agents transacting across organizational trust boundaries lack discovery, passports and certification, interoperability and attestation, plus an economics for pricing knowledge. Meanwhile vendors treat protocol presence itself as a distribution channel, publishing their capabilities as callable servers so external agent ecosystems find them without bespoke partner work.

Also: the MCP transports working group and the MCP Apps extension; Monte, a Rust sandbox, and a public dataset of observed clients; the Agent-2 framework; Nanda Town, NandaHack, a claimed 1.5B-agent deployment on India's identity and payment rails, and a City of Boston civic initiative.

Assay's desks already interoperate through a deliberately minimal medium — markdown, git refs and forge primitives — with clarification handled by filing an issue rather than by an in-band back-channel. The choose-by-workload guidance maps cleanly onto Assay's own split: desk verbs and verification rows are the strict-schema governance lane, while findings and registers are the flexible one. Clarify-instead-of-fail is already the vocabulary: question, help wanted, needs-decision.

The working group abandoned stateful in-band elicitation for exactly the reason a desk files a needs-decision issue and goes silent rather than holding an open channel: a permanently open back-channel does not survive real operations. The same evidence is a standing warning against importing live-channel thinking into any future design that treats the forge as a transport. The long-running-handoff pain point validates two shipped choices as well — durable claims on git refs and a generated board are both answers to "which context does the next agent need", and both are provenanced by construction. Where Assay is behind is that context provenance is implicit rather than a first-class field; making it explicit would strengthen any evidence-citation requirement built on top of memory.

The roster-based trust gate is a private passport and attestation scheme, and it is exactly the seam where an external attestation standard would plug in if one stabilizes — desk verbs would need signed identity to participate. That is designed and unbuilt as an interop story; there is currently no reason to federate outward. The knowledge-pricing research problem is a cousin of Assay’s own measurement problem — pricing which briefs and reviews actually paid off rather than counting activity. And protocol-as-distribution is a reminder that the public skill bundle competes in an ecosystem where discoverability increasingly runs through agent-to-agent protocols rather than documentation. A context-handoff spec is worth watching precisely because it targets the primitive Assay is built on; the counter-argument Assay holds is that markdown under review, in git, with a human gate, already supplies provenance and governance without a new spec.

• Whether the MCP transport changes ship, and whether the handshake-to-tool-call ratio improves measurably on a public server afterwards.
• Whether the minimal-envelope bet attracts a second independent implementer, or stays a standards-body artifact with one timing harness behind it.
• Whether any context-handoff spec gets adopted by a party that did not author it — the signal that markdown-plus-git is genuinely insufficient rather than merely unfashionable.
• Whether agent passports and attestation reach a draft standard a small operator could adopt without joining a consortium.