Teams and protocols: multi agent, A2A and MCP (2024 to 2025)

The first reason is the context window. One agent doing a long job fills its window with its own history, and lesson 0.9 showed what happens then. Give each sub task to a fresh agent with a clean window, and only the result comes back. The second reason is specialisation: a prompt and tool set tuned for writing SQL is worse at writing emails. The third is parallel work, several searches at once.

The research came in 2023. CAMEL in March had two agents role play a user and an assistant to finish a task. MetaGPT in August gave agents the roles of a software company, product manager, architect, engineer, with documents passed between them. Microsoft's AutoGen, also August 2023, made the general version: agents are objects that send each other messages, and any of them can be a model, a tool or a person. CrewAI and others packaged the idea for people who did not want to read papers.

In June 2025 Anthropic published how its Research feature works. A lead agent plans, spawns sub agents that search in parallel, and writes up what they return. On an internal evaluation, that system with Claude Opus 4 leading and Claude Sonnet 4 sub agents beat a single Opus 4 agent by 90.2 percent. The same post gives the bill: an agent uses about four times the tokens of a chat, and a multi agent system about fifteen times. Their advice is to use it for tasks worth that much, mostly ones where reading in parallel pays.

The field disagrees about the rest. Cognition, the company behind the Devin coding agent, published Don't Build Multi-Agents the same month. Their argument is that agents working in parallel make decisions on conflicting assumptions, the lead agent cannot see the sub agents' context, and the merged result has contradictions nobody made on purpose. Their rule is one agent and one thread of context for as long as that fits. Both positions are right about different tasks. Reading in parallel is safe. Writing in parallel is not. Part 4 of this course is about choosing.

Every team that built an agent in 2023 wrote a wrapper for GitHub, one for Slack, one for Postgres, and so did every other team. On 25 November 2024 Anthropic released the Model Context Protocol, MCP, as an open standard. A server exposes tools, each with a name, a description and a JSON Schema, plus resources and prompts. A client, the application running the model, connects to any number of servers, lists their tools and passes them to the model in the function calling format from lesson 0.8. The transport is JSON-RPC over standard input and output or over HTTP. Write the GitHub server once and every client can use it.

Uptake was fast because the function calling shape already matched. OpenAI added MCP support to its Agents SDK in March 2025, and Google's A2A announcement in April 2025 called MCP complementary and committed to both. By the second half of 2025 MCP was the default way to give a model a tool, and the open questions had moved to security: a server's tool description is text the model reads, so a malicious server is a prompt injection with a handshake.

On 9 April 2025 Google announced the Agent2Agent protocol, A2A, with more than 50 partner companies. Where MCP connects a model to tools, A2A connects an agent to other agents. Each agent publishes a card saying what it can do. A client sends it a task, gets status updates, and receives artifacts back, over HTTP. It is early, and the honest line in an interview is that MCP is in production everywhere and A2A is mostly in pilots.

Say that MCP solves a distribution problem, not a model problem: the tool is written once and reached from anywhere, in the shape the model was already trained on. Then say where that leaves the work. With tools and agents behind standard plugs, the hard part is no longer wiring. It is knowing whether the thing works, on which inputs, and what it did on the run that went wrong. That is lesson 0.12, and it is the rest of this course.