Agent protocol stack firms up
Developers are framing agent systems as a stack of reusable protocols — tool-access (MCP), agent-to-agent communication (A2A) and user-facing interfaces (AG‑UI) — instead of isolated demos, which makes integration patterns clearer for product teams. Practical guides show how to wire an MCP server to agent SDKs and swap tool access across frameworks, illustrating the plumbing engineers will reuse when building agent workflows ( ).
Agent builders are starting to sort their systems into three separate pipes: one for tools, one for agent-to-agent messages, and one for the user interface. Anthropic’s Model Context Protocol covers tool access, Google’s Agent2Agent protocol covers agent communication, and Agent User Interaction covers the front end. (anthropic.com) (developers.googleblog.com) (docs.ag-ui.com) Model Context Protocol, or MCP, was introduced by Anthropic on November 25, 2024 as an open standard for connecting assistants to repositories, business tools and development environments. On December 9, 2025, Anthropic said it was donating MCP to the Agentic AI Foundation, a Linux Foundation-directed effort co-founded by Anthropic, Block and OpenAI with support from Google, Microsoft, Amazon Web Services, Cloudflare and Bloomberg. (anthropic.com 1) (anthropic.com 2) Agent2Agent, or A2A, was announced by Google in 2025 as a way for one agent to discover and work with another agent across providers and cloud environments. Google’s recent guides describe A2A as complementary to MCP: MCP connects an agent to tools and data, while A2A handles collaboration among agents in their “natural modalities.” (developers.googleblog.com) (codelabs.developers.google.com) Agent User Interaction, or AG-UI, is the newest layer in the stack. Its public documentation says the protocol became public on April 9, 2025 and defines it as a lightweight, event-based way to connect an agent backend to chat apps, copilots and other user-facing clients. (docs.ag-ui.com 1) (docs.ag-ui.com 2) The split is becoming easier to see in practical documentation, not just architecture diagrams. Google published a “Developer’s Guide to AI Agent Protocols” in March 2026 that walks through six protocols, including MCP and A2A, as replacements for custom integration code between tools, agents and interfaces. (developers.googleblog.com) OpenAI’s Agents SDK documentation now includes first-party MCP support in both Python and JavaScript, and describes MCP as a standardized way for applications to expose tools and context to language models. The Python SDK docs say developers can plug in multiple MCP transports and reuse existing MCP servers instead of rewriting tool integrations for each framework. (openai.github.io 1) (openai.github.io 2) (github.com) That matters for product teams because the same calendar, file system or connector server can be wired into different agent runtimes with less custom glue code. Composio’s current guides show that pattern directly, including a Google Calendar MCP integration for the OpenAI Agents SDK and broader OpenAI Agents support across more than 1,000 toolkits. (composio.dev) (docs.composio.dev) (docs.composio.dev) Anthropic is also pushing MCP beyond simple tool catalogs. In a December 2025 engineering post, the company said code execution over MCP could cut context overhead by up to 98.7% when agents handle large tool sets, a sign that protocol work is moving from demos into cost and performance tuning. (anthropic.com) The result is a cleaner map of the agent stack than developers had a year ago. MCP is settling in as the tool port, A2A as the inter-agent channel, and AG-UI as the screen-facing layer that lets the rest of the system show up in products people can actually use. (anthropic.com) (developers.googleblog.com) (docs.ag-ui.com)
