Interop standards
The protocols that let agents talk to tools and to each other: a thin but load-bearing layer. MCP for tools, A2A for agent-to-agent, plus discovery and observability. Classified by interop layer.
| Protocol | Layer | Steward | Transport | Adoption | src |
|---|---|---|---|---|---|
| Model Context Protocol (MCP) | Agent ↔ tools | Anthropic | JSON-RPC 2.0 | ~1,400+ servers; adopted by OpenAI, Google, Microsoft | ↗ confirmed |
| Agent2Agent (A2A) | Agent ↔ agent | Google / Linux Foundation | HTTP / SSE | AgentCard discovery; donated to Linux Foundation 2025 | ↗ confirmed |
| Agent Communication Protocol (ACP) | Agent ↔ agent | IBM | REST | Converging into A2A | ↗ likely |
| AGNTCY | Discovery + identity | Linux Foundation | – | Open agent discovery & identity | ↗ likely |
| OpenTelemetry GenAI conventions | Observability | OpenTelemetry / CNCF | OTLP | Observability semantic conventions for agents | ↗ confirmed |
| Agent Payments Protocol (AP2) | Agent ↔ payments/commerce | Google (with FIDO Alliance) | Verifiable Digital Credentials + Mandates over A2A | Sept 2025 A2A extension; card support live; UPI/PIX/stablecoins roadmap | ↗ confirmed |
| x402 | Agent ↔ payments (HTTP-native) | Linux Foundation | HTTP 402 + stablecoins (EVM/Solana) | Blockchain-agnostic, zero protocol fees; ~75M txns / $24M / 22K sellers trailing 30d | ↗ confirmed |
| Agent Network Protocol (ANP) | Agent ↔ agent (identity + discovery) | ANP open-source community | W3C DID (did:wba) + JSON-LD | Three-tier meta-protocol for cross-agent discovery | ↗ confirmed |
| LOKA Protocol | Agent ↔ agent (identity + governance) | Academic proposal (Ranjan et al.) | Universal Agent Identity Layer (DID + VCs) | Four-layer framework; Decentralized Ethical Consensus Protocol | ↗ likely |
| Open Knowledge Format (OKF) | Agent ↔ knowledge | Google Cloud (Sam McVeety, Amir Hormati) | Markdown + YAML frontmatter (files in a folder tree) | v0.1 spec; vendor-neutral, git-hostable, one required field (type). Formalises the LLM-wiki pattern. | ↗ confirmed |
| Anthropic Agent Skills | Agent ↔ capability | Anthropic | SKILL.md + directory | Published Oct 2025 (Zhang, Lazuka, Murag) as an open format; progressive-disclosure loading (metadata → body → files); example skills at github.com/anthropics/skills | ↗ confirmed |
| AGENTS.md | Agent ↔ instruction routing | Linux Foundation (Agentic AI Foundation) | Markdown, folder hierarchy | Adopted by OpenAI Codex, Amp, Jules, Cursor and Factory; closest-file-wins routing; open spec at agents.md | ↗ confirmed |
| Cursor Rules (.cursor/rules/*.mdc) | Agent ↔ instruction routing | Cursor (Anysphere) | .mdc (Markdown + YAML) | Path-scoped conditional loading via glob patterns; sub-directory .cursor/rules/ scopes rules to monorepo packages | ↗ confirmed |
| CLAUDE.md + ~/.claude/ | Agent ↔ hierarchical memory | Anthropic | CLAUDE.md + scoped hierarchy | Claude Code memory model: scoped org → user → project → subdirectory lookup, plus an auto-memory directory the agent writes back to across sessions | ↗ confirmed |
| Cline Memory Bank | Agent ↔ prescribed memory taxonomy | Cline | Prescribed markdown taxonomy | Files: projectbrief.md, productContext.md, activeContext.md, systemPatterns.md, techContext.md, progress.md; mandatory read-at-task-start turns a stateless agent stateful purely via filesystem contract | ↗ confirmed |
| GitHub Spec-Kit (Spec-Driven Development) | Agent ↔ spec-driven workflow | GitHub | .specify/ + phase commands | Open-source toolkit at github.com/github/spec-kit; phase sequence: constitution → specify → plan → tasks → implement; each phase reads/writes its own directory | ↗ confirmed |
The thin, load-bearing layer that lets agents call tools and reach each other
Tool-calling has effectively consolidated to one protocol: MCP carries that layer and holds the adoption. Agent-to-agent has not, and is spread across four competing specifications covering identity, discovery and governance. The friction there is commercial rather than technical. A Braze agent trusting an Agentforce agent asks two vendors to extend each other a credit neither currently has much reason to give, and while native integrations exist in places, cross-vendor agent trust is not yet commonplace.
Show more
Within a single organisation the position is different. A finance agent handing work to a legal agent is a plausible near-term pattern, since both sit under one owner and one policy. The remaining barrier is confidence rather than connectivity: a receiving agent has no reliable way to judge the quality of what it has been handed. That carries a compounding risk. Where agents routinely consume other agents' output, errors propagate rather than cancel, and each layer inherits the mistakes of the one beneath it.
What is an agent protocol?
A standard for how an agent talks to something outside itself: a tool, another agent, a registry or a tracing backend. Build to the standard once and anything else that speaks it will interoperate, instead of hand-wiring every connection.
What is the Model Context Protocol (MCP)?
MCP is the standard for connecting an agent to tools and data. A server describes what a system can do, and any MCP-speaking agent can call it without a bespoke integration. It is the protocol most agent projects need first. (Model Context Protocol)
What is the difference between agent-to-tool and agent-to-agent protocols?
Agent-to-tool protocols cover one agent calling out to software, and MCP is the example. Agent-to-agent protocols cover independent agents handing work to each other, including how they find and authenticate one another. Most systems only ever need the first.
What is A2A (Agent2Agent)?
An open protocol for agents built by different vendors to discover each other and hand work between them. Google published it in 2025 and governance moved to the Linux Foundation the same year. It sits alongside MCP rather than replacing it.
Are agent protocols standardised yet?
Only the tool layer. MCP has broad adoption across model providers and software vendors. Agent-to-agent, discovery, identity and observability standards are still competing, and several are under two years old. Expect churn above the tool layer.