{
  "id": "4deb2178b18c",
  "url": "https://www.anthropic.com/engineering/built-multi-agent-research-system",
  "title": "How we built our multi-agent research system Anthropic",
  "note": "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. \n\n1. Think like your agents.\n2. Teach the orchestrator how to delegate.\n3. Scale effort to query complexity.\n4. Tool design and selection are critical.\n5. Let agents improve themselves.\n6. Start wide, then narrow down.\n7. Guide the thinking process.\n8. Parallel tool calling transforms speed and performance.\n\nThis bit out of the agents improving themselves was very interesting to me. \n\n> 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.\n\nAgent performance, collaboration, and improvement will be an interesting area as we create more of these. \n\nAlso see \n[Simon Willison's commentary](https://simonwillison.net/2025/Jun/14/multi-agent-research-system/) on this. ",
  "voice": "mixed",
  "saved": "2025-06-18T02:12:24Z",
  "host": "anthropic.com",
  "kind": "page",
  "tags": [],
  "record": "https://links.thingelstad.xyz/links/4deb2178b18c/",
  "markdown": "https://links.thingelstad.xyz/links/4deb2178b18c.md",
  "issues": [
    {
      "number": 324,
      "slug": "324",
      "title": "Weekly Thing 324 / Agents, Shortcuts, Joy",
      "date": "2025-06-21T11:00:00Z",
      "section": "Notable",
      "position": 8,
      "url": "https://weekly.thingelstad.com/archive/324/",
      "page": "https://links.thingelstad.xyz/issues/324/"
    }
  ],
  "connections": {
    "same_host": {
      "count": 30,
      "page": "https://links.thingelstad.xyz/hosts/anthropic.com/",
      "nearest": [
        {
          "id": "d8691d06cebc",
          "title": "Introducing Claude 4 Anthropic",
          "url": "https://www.anthropic.com/news/claude-4",
          "record": "https://links.thingelstad.xyz/links/d8691d06cebc/"
        },
        {
          "id": "560cdb7361bd",
          "title": "Building Effective AI Agents Anthropic",
          "url": "https://www.anthropic.com/engineering/building-effective-agents",
          "record": "https://links.thingelstad.xyz/links/560cdb7361bd/"
        },
        {
          "id": "eb9f89f8c8dd",
          "title": "Project Vend: Can Claude run a small shop? (And why does that matter?) Anthropic",
          "url": "https://www.anthropic.com/research/project-vend-1",
          "record": "https://links.thingelstad.xyz/links/eb9f89f8c8dd/"
        },
        {
          "id": "98d99aa751fb",
          "title": "Customize Claude Code with plugins Anthropic",
          "url": "https://www.anthropic.com/news/claude-code-plugins",
          "record": "https://links.thingelstad.xyz/links/98d99aa751fb/"
        },
        {
          "id": "65bde153ed6b",
          "title": "A small number of samples can poison LLMs of any size Anthropic",
          "url": "https://www.anthropic.com/research/small-samples-poison",
          "record": "https://links.thingelstad.xyz/links/65bde153ed6b/"
        }
      ]
    },
    "same_week": {
      "count": 30,
      "nearest": [
        {
          "id": "a6e53235b2f7",
          "title": "kill the metrics in your head | The Roof is on Phire",
          "url": "https://phirephoenix.com/blog/2025-05-30/metrics",
          "record": "https://links.thingelstad.xyz/links/a6e53235b2f7/"
        },
        {
          "id": "7d5ff866e555",
          "title": "POAP Product Updates - June 25",
          "url": "https://www.poap.news/product-updates-q22025/",
          "record": "https://links.thingelstad.xyz/links/7d5ff866e555/"
        },
        {
          "id": "237aa3134660",
          "title": "OpenMCP",
          "url": "https://www.open-mcp.org/",
          "record": "https://links.thingelstad.xyz/links/237aa3134660/"
        },
        {
          "id": "811d1f2b2d14",
          "title": "Writing Toy Software Is A Joy | Joshua Barretto",
          "url": "https://www.jsbarretto.com/blog/software-is-joy/",
          "record": "https://links.thingelstad.xyz/links/811d1f2b2d14/"
        },
        {
          "id": "064480f38c06",
          "title": "Field Notes From Shipping Real Code With Claude - diwank's space",
          "url": "https://diwank.space/field-notes-from-shipping-real-code-with-claude",
          "record": "https://links.thingelstad.xyz/links/064480f38c06/"
        }
      ]
    },
    "mentions": [],
    "mentioned_by": [],
    "saved_twice": []
  },
  "elsewhere": {
    "wayback": "https://web.archive.org/web/*/https://www.anthropic.com/engineering/built-multi-agent-research-system"
  }
}