Anthropic gave its AI agents three new abilities that all point the same way: agents that check their own work, fix themselves and get better over time. Here’s what they do, plus 5 simple agents you can build today.
In May 2026, Anthropic announced three new features for Claude Managed Agents, its platform for developers building agents: Dreaming, Multiagent Orchestration and Outcomes. Read together, they’re one bet. Agents that don’t just do what you say, but watch themselves, fix themselves and improve on their own.
01
Dreaming
Your brain sorts through the day while you sleep. Dreaming does that for agents. On a schedule, the agent goes back through its past sessions and memory, finds patterns (mistakes it keeps making, steps that work, preferences your team shares) and rewrites its own notes. The model itself doesn’t change. Its playbook does, so tomorrow it starts smarter.
Status: research preview. Developers request access, and you choose whether memory updates apply automatically or wait for your review.
02
Multiagent Orchestration
A lead agent breaks a big job into pieces and hands each piece to a specialist agent with its own model, instructions and tools. The specialists work at the same time on shared files, and the lead pulls it all together. Think a manager with a team instead of one person doing everything.
Status: public beta. In October, Anthropic added a version where the agent writes and runs these team workflows itself.
03
Outcomes
You write a rubric for what "good" looks like. The agent does the work, and a separate grader checks it against your rubric. If it falls short, the agent keeps revising until it passes. No more "first draft and stop."
Status: public beta.
All three live on Anthropic’s developer platform, not in the regular Claude app, so you probably won’t flip them on yourself. (Latest release notes) But the direction is clear:
The people building simple agents now will know exactly how to use these when they show up in everyday tools.
You don’t need developer features. Each of these works in the Claude app on a paid plan. Connect your tools under Customize, then Connectors. Paste the prompt into a chat and Claude sets it up as a scheduled task.
01
Inbox Triage
Every morning it finds the emails that actually need you and drafts replies for you to review.
02
Weekly Numbers Report
Pick one tool you care about, like Shopify, Stripe or Google Analytics, and get the same clear report every Monday.
03
Daily Focus Brief
Reads your calendar and priorities and tells you what to focus on today.
04
Content Idea Engine
Looks at what’s working in your niche and gives you 5 specific ideas every week.
05
Customer Voice Miner
Reads your reviews and support emails and tells you what customers love, what they complain about and the words they use.
Want to build a real agent with context files and memory? Here’s how I build all of mine.