Bottom line
Inference, low confidence: if your agents must discover and delegate to agents from other teams, target A2A v1.x through one of its official SDKs, and do not design your own protocol. Wrap it: pin the shared compatibility test kit (the TCK) and run it against your build, test every transport the spec requires, and call any interop result "self-interop" until you have tested against an implementation from outside the A2A organization, because the evidence found none. Confidence is low because the evidence pack itself calls the category thin.
Adopt, do not rebuild
License checked on 2026-09-23; non-permissive licenses are named in the text.
- a2aproject/A2A Apache-2.0: the spec, now under vendor-neutral stewardship with tagged v1.0.0 and v1.0.1 releases; IBM's competing ACP was archived after merging into it, so a new project should target A2A and treat ACP as history. [Verified] (ecosystem/pickup/_evidence/multi-agent-protocols.md:8
Tagged releases v1.0.0 (2026-03-12) and v1.0.1 (2026-05-28)
; ecosystem/pickup/_evidence/multi-agent-protocols.md:17Archived Aug 2025 after merging into A2A
) - a2aproject/a2a-python, a2a-js, a2a-go Apache-2.0: official SDKs whose test trees were read (compat suites per spec version, a wire-format corpus diffed between two SDKs, per-transport tests, TCK wiring); writing a fourth implementation of the same wire would recreate them. [Verified] (ecosystem/pickup/_evidence/multi-agent-protocols.md:9
Official Python SDK for A2A — pushed today
; ecosystem/pickup/_evidence/multi-agent-protocols.md:22found by round-tripping a shared corpus through this SDK and a2a-python
; ecosystem/pickup/_evidence/multi-agent-protocols.md:21e2e/tck/run_tck.sh, e2e/tck/orchestrate_tck.py
) - agent-network-protocol/anp Apache-2.0: the one non-A2A protocol with a spec community and a reference implementation, for builders who need decentralized identity (W3C DID) and end-to-end encryption between agents; it ships a large message-layer unit suite. [Verified] (ecosystem/pickup/_evidence/multi-agent-protocols.md:13
three-layer stack (W3C DID did:wba identity + encrypted comms
; ecosystem/pickup/_evidence/multi-agent-protocols.md:27agent-network-protocol/anp:anp/unittest/ (74 files)
)
Copy these practices
- Run the standalone TCK against your implementation: keep the conformance kit in its own repository, clone it at a pinned commit, and run it against a local system under test, so every implementation meets the same bar. Starter kit: Incumbent oracle pinned before implementation. [Verified] (ecosystem/pickup/_evidence/multi-agent-protocols.md:21
keep the conformance suite in its own repo so every implementation (and a clean-room rewrite) tests against the same bar
) - Known-broken cases as tests that must flip: record a known wire bug as a test expected to fail, so fixing the bug forces the marker to be updated in the same change. Starter kit: Every falsified hypothesis banked with a retry predicate. [Verified] (ecosystem/pickup/_evidence/multi-agent-protocols.md:22
track known failures as tests that must flip, not as skipped TODOs
) - Same behaviors on every binding: when the spec requires JSON-RPC, gRPC and REST, one test plan runs on all of them; one green binding is not conformance. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/multi-agent-protocols.md:24
if the protocol has multiple bindings, test the same behaviors on every binding, not just one
) - Compat suites per old spec version: keep a test directory per earlier spec version so new servers keep serving old clients. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/multi-agent-protocols.md:23
keep per-spec-version compat test dirs so vN servers keep working with vN-1 clients
) - Test signing and auth as their own specs: signed agent cards and auth schemes get dedicated test files, and handshake, signing and discovery are unit-tested with no model in the loop. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/multi-agent-protocols.md:28
security features (signing, auth schemes) get their own spec files
; ecosystem/pickup/_evidence/multi-agent-protocols.md:27unit-test the handshake/crypto/discovery primitives independently of any LLM
)
Build only if
- No evidenced constraint justifies a new protocol: the category consolidated from two specs to one, and the companion's own charter rules out new protocol design. Building a clean-room implementation of A2A is defensible only where no official SDK exists for your language, and even then the TCK is the bar. Whatever you build or adopt, count a result as full interop only against a peer implementation from outside the originating organization and call anything else self-interop; that is the companion's rule, and the interop tested in this category is inside the A2A organization's own SDKs. [Inference] (ecosystem/pickup/_evidence/multi-agent-protocols.md:33
A new clean-room project should treat A2A v1.x as the target and ACP as historical
; ecosystem/pickup/pickup-multi-agent-protocols.md:21New protocol design (the charter is conformance, not invention)
; ecosystem/pickup/pickup-multi-agent-protocols.md:44full INTEROP requires ≥1 independent peer implementation
; ecosystem/pickup/_evidence/multi-agent-protocols.md:32There is no second fully independent A2A implementation outside the a2aproject org found in this sweep
)
Where FrankenSuite touches this
- None of the 44 packets implements an agent-to-agent protocol. The nearest is frankenterm's Robot Mode, a typed JSON envelope with golden test matrices for one agent driving others through terminal panes, which its packet proposes testing against a non-frankenterm backend before calling it a standard. [Verified] (packets/frankenterm-assessment.md:37
is the closest thing in the wild to an agent-to-terminal control contract
) - franken_alignment models commit-reveal voting rounds among helper agents as an in-memory reference model only; its packet grades it TRL 3 and nothing production exists. [Verified] (packets/franken_alignment-assessment.md:19
salted commit–reveal helper congresses
; packets/franken_alignment-assessment.md:196Technology readiness | TRL 3
) - The suite's own multi-agent development coordinates through MCP Agent Mail, a development tool the packets describe but did not assess as a product. [Maintainer claim] (packets/franken_nlp-assessment.md:67
MCP Agent Mail coordination, bead-assignee locking
)
What we cannot say
- Whether A2A interop holds across organizations: the category is thin, and the interop tested is inside the A2A organization's own SDK matrix (ecosystem/pickup/_evidence/multi-agent-protocols.md:32
Thin category, honest assessment.
; ecosystem/pickup/_evidence/multi-agent-protocols.md:32There is no second fully independent A2A implementation outside the a2aproject org found in this sweep
). - Whether passing the TCK means full conformance: the kit is young and no evidence shows all four SDKs run it in CI (ecosystem/pickup/_evidence/multi-agent-protocols.md:34
no evidence yet that all four SDKs run it in CI
). - ANP's cross-language interop: its named interop workflow covers Rust and Python only, not all five languages (ecosystem/pickup/_evidence/multi-agent-protocols.md:32
covers only rust-python, not the full five-language matrix
). - The .NET SDK was checked for activity only, so it is not listed as an incumbent above (ecosystem/pickup/_evidence/multi-agent-protocols.md:12
Official .NET SDK, pushed yesterday
). - Coral Protocol was excluded without an API check, and MCP is out of scope because it connects agents to tools, not agents to agents (ecosystem/pickup/_evidence/multi-agent-protocols.md:35
not verified via API and not cited per the no-memory rule
; ecosystem/pickup/_evidence/multi-agent-protocols.md:36Model Context Protocol is agent-to-tool, not agent-to-agent
). - Stars and push dates show the SDKs are alive, not that they interoperate.
Revisit when
- An A2A implementation from outside the a2aproject organization appears and passes the pinned TCK: interop claims can then move past self-interop, and confidence can rise.
- All four official SDKs are observed running the TCK in CI.
- A2A publishes a v2 or a breaking v1.x revision: re-pin the TCK and the compat suites before any claim.
Sources
- Pickup companion: ecosystem/pickup/pickup-multi-agent-protocols.md, the charter seed, oracle candidates, and first claim registry for this area.
- Evidence pack: ecosystem/pickup/_evidence/multi-agent-protocols.md, the repositories checked against the GitHub API, with dates.
- Verdict source: stack/multi-agent-protocols.md, written under stack/METHOD.md.
Starting a build here?
Begin from the pickup companion for Agent-to-agent protocols (A2A class): it carries the charter seed, the oracle candidates and how to verify them, and the first claims to register, so a new project starts from evidence instead of from scratch.