01What Claude AI Agents Are and How Dreaming Works

Claude AI agents are Anthropic's AI systems that can run tasks for you, like handling emails or writing reports. The new "Dreaming" feature, announced at the Code with Claude event in San Francisco, lets these agents review their past work while they are idle. This is like a human dreaming at night, but for AI.

During a dream, the agent goes through up to 100 of its past chats. It looks at what worked and what failed. It spots patterns you could never see.

Then it cleans up its memory, removes duplicate notes, and fixes things that don't match. It locks in lessons. When you wake it up, it's a smarter version of itself.

This means the agent fixes itself while you sleep, eat, or spend time with your kids.

02Key Features: Dreaming, Outcomes, Multi-Agent, Webhooks

Anthropic dropped four big updates at the event. Dreaming is the headline feature, but the others are just as important.

  • Outcomes (public beta): You write a checklist of what a good result looks like. A second, separate Claude grades the agent's work against that checklist. If it fails, the grader points out what's wrong, and the agent tries again until it passes.
  • Multi-agent orchestration (public beta 2): You set up one lead agent. It breaks a big job into smaller parts and hands each part to a specialist agent. Up to 20 agents can work on one project at the same time, sharing files and checking in with each other.
  • Webhooks: You set an outcome, walk away, and get a ping when the job is done. Like a text from a co-worker that just says "done."

All four features work together. The agent learns (Dreaming), grades itself (Outcomes), splits jobs across a team (Multi-agent), and tells you when it's done (Webhooks).

03How to Set Up Claude AI Agents With Dreaming

To get started with Claude AI agents and the new features, follow these steps. Anthropic didn't give a detailed setup guide in the episode, but the basics are clear.

  1. Open the Claude console (the place where you manage your agents).
  2. Create a new agent or choose an existing one.
  3. Enable the Dreaming feature if it's available in your plan. It's in research preview right now.
  4. For Outcomes, write a checklist of what a good result looks like. For example, a newsletter rubric: must have a strong hook, must have one customer story, must end with one clear call to action.
  5. For Multi-agent orchestration, set up a lead agent and then define specialist agents with their own tools and jobs.
  6. Use webhooks to connect your agent to your CRM, email tool, or Slack, so you get notified when tasks finish.

Once set up, the agent runs in the background. You can go live your life.

04How to Use Claude AI Agents for Real Tasks

Let's walk through examples from the episode. Imagine an agent that handles customer emails. Day one, it makes a small mistake, like using the wrong tone.

Day two, same thing. Old agents would keep doing this forever. But with Dreaming, the agent reviews those emails between sessions.

It sees the pattern and writes a new note: "Use a friendly tone, not a formal one." Next session, it's fixed. You didn't say a word.

Another example: you want a weekly newsletter. You write a rubric (Outcomes). The agent writes.

The grader reads. If it misses the hook, it gets sent back. If it misses the story, sent back.

When you wake up, the newsletter is done and right, first try from your point of view.

For multi-agent orchestration, use it for big projects. A lead agent takes the order, asks follow-up questions, and sends the actual writing to bigger Opus agents. If you ask for three drafts, three Opus agents work on them at the same time.

You get three options in the time it takes to do one.

05Results and Tests: Harvey's 6x Completion Rate

Anthropic shared test results that show the impact. Harvey, a legal AI firm, said their agent task completion rates went up six times after using Dreaming. That's not a small lift. It's the difference between an agent that costs you time and one that gives you time back.

Anthropic's own test on Outcomes showed task success went up by 10 points. The hardest tasks saw the biggest jump. Outputs like Word docs went up 8%, and slide decks went up 10%.

So if you've ever asked an agent to make a sales deck and got back something off, this fixes that.

For multi-agent orchestration, the episode gave an example with a writing tool called Spiral from a company called Every. Their lead agent runs on Haiku, the fast model. It takes orders and asks follow-up questions.

It sends the actual writing to bigger Opus agents. With three drafts, three Opus agents work on them at the same time. You get three options in the time it takes to do one.

06Pricing, Limits, and Mistakes to Avoid

Pricing for managed agents is $0.08 per agent runtime hour, on top of normal model cost. Dreaming, Outcomes, and Webhooks are included in that rate—no extra charge. You're not paying more for the agent to grade itself or learn from past sessions.

It just does it.

One limit to know: Dreaming is in research preview right now, not fully public. Outcomes is in public beta, and multi-agent orchestration is in public beta 2. So some features may change.

Avoid these mistakes mentioned in the episode:

  • Don't treat agents like a fancy chat tool. They now run jobs in the background.
  • Don't ignore the traceability feature. You can pop open the console and see exactly which agent did what, in what order, and why. Use it to find problems.
  • Don't assume agents forget mid-job. Events are persistent, so they remember where they left off. You can run jobs that take hours or days.

07Why Claude AI Agents Outperform OpenClaw

The episode compares Claude's new agents to OpenClaw. OpenClaw is built around the idea of an agent in a browser. It's good at clicking, scrolling, and filling out forms.

But it doesn't dream, it doesn't grade itself, and it works one task at a time.

Claude just walked past OpenClaw on the curve that actually matters: can the agent learn and run jobs without you babysitting? Right now, Claude can. OpenClaw can't.

Their agent tools can't yet either.

If you've been pushing OpenClaw to run your business, you've been hitting the same wall over and over. Claude jumped that wall. The agent that forgot your rules now keeps them.

The agent that got tone wrong now grades its own tone. The agent that took forever now splits the job across 20 helpers.

08Who Claude AI Agents Are For and Why It Matters

These agents are for business owners who want to save time and put follow-ups, content, lead gen, and customer support on autopilot. If you tried agents six months ago and gave up because they were forgetful, this is your sign to look again. The thing that made you quit just got fixed.

Why does this matter? Because agents now learn, grade themselves, split jobs, and report back. Once an agent has been dreaming for 30 days, it's a totally different worker than when it started.

Imagine an employee that gets 6% better every week without you doing anything. That compounds fast.

The folks who set this up now are about to leave everyone else behind—not in a year, but in a few months. The whole multi-agent setup is fully traceable, so you can find problems fast. And you can plug it into your CRM, email, or Slack with webhooks.

Agents run quiet until you're needed.