01What Claude Managed Agents Are and How They Work

Claude managed agents are AI systems from Anthropic that can handle complex tasks, learn from experience, and connect to your tools. In this episode, Julian Goldie breaks down four new features that make these agents work like a team of improving employees.

The agents log every task, decision, and problem. Then, between sessions, the new Dreaming feature organizes that information into memory. The agents use what they learn to get better over time.

You stay in control. You can let Dreaming update memory automatically, or you can review changes before they go live. This makes managed agents useful for businesses that want AI to improve without constant babysitting.

02The Four Big New Features in Claude Managed Agents

Anthropic announced four additions to Claude managed agents. Each one helps in a different way, and together they change how agents work.

  • Dreaming: A scheduled process that reviews past sessions and memory stores. It finds patterns, fixes mistakes, and updates memory so agents improve.
  • Outcomes: A separate grading agent checks outputs against a rubric you write. If the output fails, the grader sends feedback and the agent retries automatically.
  • Multi-Agent Orchestration: A lead agent breaks a big job into parts and delegates each part to specialists. These specialists work in parallel with their own models and tools.
  • Webhooks: Agents can trigger external apps and receive events. This connects Claude to your CRM, email, project management, and more.

Each feature solves a common problem. Dreaming stops agents from starting over. Outcomes cuts your review time.

Multi-agent speeds up big tasks. Webhooks automate handoffs.

03How to Get Started with Claude Managed Agents

Getting started is simple. Julian suggests you start small and build habits. Here is a step-by-step plan from the episode.

  1. Pick one repeating workflow, like writing daily content or answering customer inquiries.
  2. Write a basic rubric. List what good output looks like, such as tone, length, and required steps.
  3. Set up an outcomes loop. The grader checks every draft against your rubric.
  4. Review the results for a few days. See what the grader catches.
  5. If the workflow works, try adding Dreaming. Request access to the research preview.
  6. For bigger projects, map out which steps could run in parallel. Then build a multi-agent system.
  7. Finally, add webhooks to connect the agent to your tools like CRM or project management.

You do not need multi-agent on day one. Start with outcomes and build from there.

04How to Use Claude Managed Agents in Your Business

Julian shows practical examples from the episode. For content creation, you write a rubric that says: tone must match your voice, include one clear action step, stay under 150 words, and avoid jargon. The grading agent checks every draft against those rules.

If something is off, the agent fixes it before you see it.

For client onboarding, a webhook can fire when the agent finishes. The client's details go straight into your CRM, a welcome email starts, and a task gets created in your project management tool. No copy-pasting.

For legal work, Harvey uses managed agents with Dreaming. Their agents remember file type workarounds and tool patterns between sessions. That improves long-form drafting and document creation.

The key is to think of your workflow as a team. A lead agent manages specialists. Each specialist improves from past experience.

Every output gets checked before it leaves the system.

05Results and Tests Shown in the Episode

The episode shares real numbers from Anthropic's testing and from Harvey. These show the features work.

  • Harvey's completion rates went up around six times with Dreaming.
  • Outcomes improved task success by up to 10 percentage points over a standard prompt approach. The biggest gains were on harder tasks.
  • Generation quality improved 8.4% for Word documents and 10.1% for presentations with Outcomes.

These numbers come from internal tests and early users. They show that agents get better with feedback and memory. Even small gains add up over time.

06Tips and Limits for Claude Managed Agents

Julian shares several tips and warnings from the episode. First, Dreaming is still in research preview. You have to request access.

It is not available to everyone yet.

Second, start simple. Do not build multi-agent systems on day one. Find one workflow and set up an outcomes loop first.

Third, use a rubric that is specific. Vague rubrics lead to fuzzy grading. Be clear about tone, length, and required elements.

Fourth, you stay in control. You can review Dreaming updates before they go live. You do not have to let it change memory automatically.

Finally, webhooks are quiet but powerful. They connect Claude to your real-world tools. If you skip them, you lose the automation benefits.

07Who Should Use Claude Managed Agents and Why It Matters

This update is for anyone who runs a business and uses AI. If you write content, do research, or handle client communication, you can benefit.

Julian explains that the first session might be average, the tenth better, and the hundredth tailored to your business. This changes the economics of AI. You get compounding improvement without retraining.

The episode quotes Winston Weinberg, CEO of Harvey, who says the legal industry is already past AI as an assistant and into the era of AI agents. That shift is happening in every industry.

The four features are the infrastructure for that era. They are out now and working. Companies are already seeing results.

The question is whether you build the systems now or play catch-up later.