JUST IN: Claude Managed Agents now support multiagent orchestration, an outcomes loop for rubric-driven self-improvement, dreaming for self-learning, & webhooks 😳

Here’s why each of these is really powerful:

🤖 Multi-agent orchestration = Teamwork mode

The main AI can now split a big job into smaller pieces and hand them off to specialized "helper AIs" (sub-agents). Example: One helper reviews code, another tests it, another writes docs. They all share the same files and workspace so they don’t get confused and can work together smoothly.

🤖 Outcomes loop (rubric-driven self-improvement) = The AI grades its own homework

You give it a clear checklist of what "good enough" looks like (called a rubric). A separate "grader" AI checks the work and says “not quite - try again.” The agent keeps improving until it passes. This means the AI actually gets better at your specific task the more it does it.

🤖 Dreaming = Background self-learning

While the AI isn’t working on your stuff, it quietly reviews everything it did in past sessions. It organizes memories, spots patterns, adds new skills, and deletes old/useless info. Basically, the AI is “sleeping on it” and waking up smarter each time.

🤖 Webhooks = Instant “I’m done!”notifications

Instead of you constantly checking or keeping a connection open, the agent can just ping your app or computer when it finishes (or hits any milestone). Super convenient for real apps.

These changes turn Claude’s agents from cool demo tools into something that can handle real, complex, long-running work on its own, and actually improve itself over time without you babysitting it.