How we built our multi-agent research system Anthropic

As we build Agents we need to use a mental model of agents performing tasks that are relatively clear. We can do complicated things by orchestrating these tasks together. That is a multi-agent system with multiple agents coordinating. This article shares how Anthropic is approaching these.

1. Think like your agents.
2. Teach the orchestrator how to delegate.
3. Scale effort to query complexity.
4. Tool design and selection are critical.
5. Let agents improve themselves.
6. Start wide, then narrow down.
7. Guide the thinking process.
8. Parallel tool calling transforms speed and performance.

This bit out of the agents improving themselves was very interesting to me.

We found that the Claude 4 models can be excellent prompt engineers. When given a prompt and a failure mode, they are able to diagnose why the agent is failing and suggest improvements. We even created a tool-testing agent--when given a flawed MCP tool, it attempts to use the tool and then rewrites the tool description to avoid failures. By testing the tool dozens of times, this agent found key nuances and bugs.

Agent performance, collaboration, and improvement will be an interesting area as we create more of these.

Also see Simon Willison's commentary on this.