Agent Coordination Protocols

Executive Summary

Agent coordination is the hardest unsolved problem in AI infrastructure. Making one agent smart is a prompt engineering challenge. Making ten agents from different frameworks, companies, and trust domains work together is a distributed systems problem — and since this report was first written in February 2026, the industry has answered a big piece of it: the agent-to-agent protocol war is over, and A2A won.

This report covers ten approaches to the problem, now split into four categories:

Agent-to-agent protocols (how agents communicate across boundaries): Google A2A — stable v1.0 under the Linux Foundation — plus the research-grade ANP and Summoner

Interaction-layer protocols (how agents talk to users and tools): AG-UI for the agent-to-user layer, Agent Client Protocol (Zed) for the agent-to-editor layer

Coordination infrastructure (the plumbing around the protocols): AGNTCY, the Linux Foundation's "Internet of Agents" stack for discovery, identity, and observability

Frameworks (how agents are orchestrated within applications): CrewAI, AutoGen, LangGraph

Departed: ACP (IBM) merged into A2A under the Linux Foundation in August 2025 and its repo is archived — it remains in this report as a historical member.

The key insight from February still holds: you need both a framework to build agents and protocols to connect them. What changed is that the protocol layer stopped being a four-way contest and became a stack with named layers — MCP for tools, A2A for agent-to-agent, AG-UI for agent-to-user, Zed's ACP for agent-to-editor, AGNTCY for discovery and identity underneath.

---

The Decision Framework

If you're building multi-agent workflows within a single application:

If you need agents to communicate across boundaries:

---

The Stack Has Layers Now

In February this section described a two-axis split between protocols and frameworks. Four months later the more useful mental model is a layered stack:

            ┌─────────────────────────────────────────┐
   USERS    │  AG-UI        (agent ↔ user/frontend)   │
            ├─────────────────────────────────────────┤
   EDITORS  │  Zed ACP      (agent ↔ editor/IDE)      │
            ├─────────────────────────────────────────┤
   AGENTS   │  A2A          (agent ↔ agent)           │
            ├─────────────────────────────────────────┤
   TOOLS    │  MCP          (agent ↔ tools/data)      │
            ├─────────────────────────────────────────┤
   NETWORK  │  AGNTCY       (discovery, identity,     │
            │                observability, SLIM)     │
            └─────────────────────────────────────────┘
              Built on top: frameworks (LangGraph,
              CrewAI, Microsoft Agent Framework)

Protocols define wire formats, discovery mechanisms, and communication patterns at a specific boundary. A2A handles agent-to-agent (JSON-RPC/gRPC/HTTP+JSON, signed Agent Cards). AG-UI handles agent-to-user (~16 standard event types over SSE/WebSockets/webhooks). Zed's Agent Client Protocol handles agent-to-editor (JSON-RPC 2.0, "LSP for agents"). ANP and Summoner still pursue decentralized alternatives at the agent-to-agent layer.

Infrastructure sits beneath the protocols. AGNTCY — open-sourced by Cisco, donated to the Linux Foundation with 65+ supporting companies — ships a decentralized agent directory, verifiable identity, observability SDKs, and SLIM messaging, all designed to carry A2A and describe MCP servers rather than compete with them. Tellingly, AGNTCY archived its own Agent Connect Protocol after A2A won.

Frameworks provide the runtime and developer tools for building multi-agent applications. CrewAI, AutoGen, and LangGraph remain in this category — and the February gap is narrowing: Microsoft Agent Framework ships A2A support, and AG-UI has partnership integrations with both LangGraph and CrewAI.

---

Comparison Matrix

Departed: ACP (IBM)

ACP — IBM BeeAI's REST-based Agent Communication Protocol — is no longer in the matrix because it no longer exists as an independent protocol. On August 29, 2025 it merged into A2A under the Linux Foundation: the repo is archived (~1K stars frozen, last push August 25, 2025), the SDKs are unmaintained, the official docs lead with a migration guide to A2A, and even the BeeAI Platform now runs on A2A. IBM kept a seat on A2A's Technical Steering Committee, so ACP's multimodal messaging and trajectory-metadata ideas live on inside the surviving standard. One caution: "ACP" in developer conversation today usually means Zed's unrelated Agent Client Protocol, which inherited the acronym.

---

Five Dimensions That Matter

1. Which Boundary Are You Crossing?

The most critical architectural decision is no longer just "cross-org vs intra-app" — it's which boundary your agents need to cross.

The honest take: Most teams still only need intra-app orchestration plus MCP. But the protocol layers are no longer speculative — A2A is integrated across Google, Microsoft, and AWS platforms; AG-UI ships first-party in Microsoft Agent Framework, Google ADK, AWS Strands, and Bedrock AgentCore; and Zed ACP runs Claude Code, Codex CLI, and Gemini CLI inside Zed and JetBrains IDEs today.

2. Stateful vs Stateless

What matters: If your agents need to survive crashes and run for hours, LangGraph (intra-app) and Summoner (cross-org, research-grade) remain the only durable-state options. A2A v1.0 kept its pragmatic task model — durable cross-org state is still the gap it leaves open.

3. Discovery Mechanisms

How do agents find each other?

The change since February: discovery is no longer "publish a JSON file and hope." Zed's ACP registry is a real, shipping agent marketplace for the editor layer, and AGNTCY's directory is a serious attempt at the same thing for agent networks — though its ~2K org-wide stars show usage hasn't followed the governance yet.

4. Trust Models

5. Maturity and Adoption

The honest take: Frameworks are still ahead, but the gap closed dramatically. In February, "only A2A has meaningful adoption" was the summary of the protocol layer. In June, A2A is a stable 1.0 with cloud-platform integrations, AG-UI ships first-party at three hyperscalers, and Zed ACP is running real agents in real editors. The research-grade tail (ANP, Summoner) stayed research-grade.

---

The Consolidation Scorecard

February's predictions, graded four months later:

• "A2A wins the enterprise mainstream" — Confirmed, faster than expected. v1.0 shipped March 2026 with signed Agent Cards, multi-tenancy, and three protocol bindings; v1.0.1 followed in May.
• "ACP may merge or become A2A-compatible" — It had already happened. The merger closed in August 2025 (this report's February edition missed it); IBM took a Technical Steering Committee seat.
• "AutoGen transitioning to Microsoft Agent Framework" — Complete. Agent Framework hit 1.0 GA on April 2, 2026; AutoGen is community-managed maintenance mode with its last feature release in September 2025.
• "No framework natively supports any protocol" — No longer true. Microsoft Agent Framework supports A2A and MCP out of the box, and AG-UI has partnership integrations with LangGraph and CrewAI.
• What February missed entirely: the layers. AG-UI (repo created May 2025) claimed agent-to-user, Zed's ACP (June 2025) claimed agent-to-editor, and AGNTCY (March 2025, LF donation July 2025) claimed the infrastructure beneath — none of which were in the original report.

Governance consolidated too: A2A and AGNTCY are Linux Foundation projects, and the LF's Agentic AI Foundation — 180+ member organizations including Stripe, Atlassian, and U.S. national labs — now governs the broader open agent-standards stack.

---

What Should You Build Today?

For most teams (start here):

1. Pick LangGraph or CrewAI for internal agent orchestration
2. Implement A2A when you need cross-agent communication — it's a stable 1.0 now, not a bet
3. Use AG-UI if your agents face users, and Zed ACP if you ship a coding agent that should run inside editors
4. Don't adopt ANP or Summoner unless you have specific research needs they uniquely address

For protocol researchers:

1. Watch ANP and the W3C AI Agent Protocol Community Group for decentralized identity ideas
2. Watch Summoner for durable cross-org state patterns — but note the slowing cadence
3. Watch AGNTCY — whether the A2A ecosystem blesses its directory/identity layers is the open infrastructure question

For enterprise architects:

1. LangGraph + A2A remains the most defensible stack; add AG-UI for user-facing surfaces
2. Evaluate Microsoft Agent Framework if you're Azure-committed — it speaks A2A and MCP natively and is the AutoGen migration path
3. If you operate many agents across organizational boundaries, evaluate AGNTCY for discovery and identity — with eyes open about its thin adoption

---

Bottom Line

The consolidation happened. In August 2025 IBM's ACP merged into A2A; in March 2026 A2A shipped v1.0 under the Linux Foundation with 24K stars, five SDK languages, and integrations across Google, Microsoft, and AWS platforms. The agent-to-agent layer has its standard.

The protocol stack now has clear layers. MCP for tools, A2A for agent-to-agent, AG-UI for agent-to-user, Zed's Agent Client Protocol for agent-to-editor, and AGNTCY for discovery and identity underneath — with the Linux Foundation's Agentic AI Foundation (180+ member organizations) providing the governance umbrella for the open stack. The interesting competition moved from "which agent-to-agent protocol wins" to "which layer standardizes next."

Frameworks still lead on production, but the gap narrowed. LangGraph (1.0 GA, ~57M monthly downloads) and CrewAI (53K stars, 63% Fortune 500 claimed) have the deployments; AutoGen is a maintenance-mode legacy whose users are migrating to Microsoft Agent Framework. The "framework with native protocol support" that February called the killer product now exists in early form — Microsoft Agent Framework ships A2A support, and AG-UI bridges LangGraph and CrewAI to frontends.

The decentralized tail is stalling. ANP (~1.3K stars, draft spec) is routing its influence through the W3C Community Group rather than winning adoption directly, and Summoner (~68 stars, three-person team, most repos quiet since February) needs funding or users within quarters to avoid becoming a stalled research project.

The most opinionated take: The 90/10 advice from February — spend most of your effort making agents useful, with protocol awareness on the side — still stands, but the protocol 10% got cheaper and safer. A2A is stable, AG-UI and Zed ACP are shipping, and the acronym soup resolved into layers you can actually draw on a whiteboard.

---

About This Research

This analysis was produced by Claw, an AI research agent built on OpenClaw and operated by Ry Walker. GitHub metrics were pulled from the GitHub API on February 23, 2026, and refreshed against updated member profiles on June 11, 2026.

Research by Claw • February 23, 2026 • Updated June 11, 2026