01What the Claude managed agents update is

The Claude managed agents update is a set of four new features from Anthropic that make Claude agents smarter, faster, and more reliable. The headline feature is Dreaming, which lets the AI review its own past work, find mistakes, and improve on its own.

The other three features are Outcomes, multi-agent orchestration, and webhooks. Together, they turn Claude from a simple assistant into a self-improving team that can work in the background while you focus on other things.

02Four big upgrades in the update

Anthropic announced these features at the Code with Claude event on May 6th. Dreaming is in research preview, while Outcomes, multi-agent orchestration, and webhooks are in public beta.

  • Dreaming: A background process that reviews past sessions, cleans up messy memory, and writes new notes for the agent to use next time.
  • Outcomes: A separate grader agent checks work against a checklist you write. If the work fails the bar, the main agent tries again.
  • Multi-agent orchestration: A lead agent breaks a big job into pieces and hands each to a specialist agent. They work in parallel and share files.
  • Webhooks: Built-in notifications let you know when an agent finishes a job, via Slack, email, or your own app.

03How to set up the Claude managed agents update

You can request access to these features today. Developers can use them through the Claude managed agents docs.

  1. Go to the managed agents documentation on Anthropic's site.
  2. Request access for the features you want, like Dreaming or Outcomes.
  3. Pick a task you do regularly, like customer replies or weekly reports.
  4. For Outcomes, write a checklist of what a good result looks like.
  5. Set up a starting prompt for your agent.
  6. If you use Dreaming, schedule a background session between your work sessions.
  7. Test the setup with a small job before running it on important tasks.

04How to use Dreaming, Outcomes, and the rest

Dreaming is simple. When your agent finishes a task, it writes notes to memory. Over time, that memory gets messy with duplicates and old info.

Dreaming runs between your sessions, reads all past work, pulls out patterns, and cleans up the memory. It writes new notes the agent can use next time.

Anthropic showed an example with six landing sites for a space mission. The first run was good but not perfect. They pressed one button to trigger a dreaming session.

Overnight, the agent reviewed past sessions and wrote a playbook. The next morning, the new run was much better, especially on the sites that failed before.

With Outcomes, you write a checklist for what good looks like. A separate grader agent checks the final result against that list. If it fails, the grader says what's wrong and the main agent tries again.

The grader runs in its own fresh context, so it isn't biased by the agent's thinking.

Multi-agent orchestration works like a team. A lead agent breaks a job into smaller pieces and hands them to specialist agents. Each specialist has its own model, prompt, and tools.

They all work at the same time and share notes. The lead agent checks the mid-job because everything is saved.

Webhooks let one app ping another when something happens. You start a Claude agent, walk away, and get a notification when it finishes. That means you don't have to sit and wait or keep checking.

05Results and tests from the episode

Harvey, a legal AI company, tested Dreaming with their agents. They handle long documents and legal drafting. With Dreaming on, their agents remembered small things like file format quirks and tool-specific tricks.

Task completion rates went up around six times.

That wasn't from a smarter model. It was from the agent carrying lessons across sessions. WiseDocs, a medical document review company, built a quality check agent using Outcomes.

They cut document review time by 50%.

Netflix also processed logs they could never get through manually. These results show the update can make a big difference right away.

06Tips, limits, and mistakes to avoid

The biggest mistake is doing nothing. Most people will watch this update and not try it, then wonder why competitors are ahead in six months. Don't be that person.

  • Start with Outcomes: It's the easiest feature to try first. Pick one weekly task, write a rubric, run it through Outcomes, and see the difference.
  • Pick one process: Choose a task that eats up hours every week, like customer replies or lead follow-up, and make that your first agent.
  • You need a clear job: You don't need to be a coder, but you need a clear job, a checklist for what good looks like, and a starting prompt.
  • Check the docs: Use the managed agents docs for setup details.

One limit is that Dreaming is in research preview, so it may not be available to everyone yet. The other features are in public beta, so they're live but still new.

07Who this update is for and why it matters

This update is for small business owners, marketers, coaches, and service providers. If you already use Claude, you can now use it like a team instead of a smart assistant. You give it a goal, the team runs in the background, checks each other's work, and gets smarter every week.

The key shift is from one chat at a time to agents running your repeat tasks while you sleep. This matters because the businesses using these features are getting results in months that used to take years. Learning to use this in the next 90 days can set you apart from those who don't.