Multi‑agent orchestration arrives in Claude Managed Agents public beta

AI agents are moving from “one model, one task” into something closer to a small software team. That matters because the hard part is no longer just getting a model to answer well — it’s g(thenewstack.io) and split jobs across parallel workers. That gap is exactly what Anthropic is trying to close. On May 6, it expanded Claude Managed Agents with dream(claude.com) public beta it launched on April 8. (thenewstack.io) ### What is Man(claude.com)rvice for running long-horizon agents on its own infrastructure. Instead of developers wiring together sandboxes, state storage, tool execution, tracing, and recovery logic themselves, they define the task, tools, and guardrails and Anthropic runs the harness. Anthropic’s pitch is basically “focus on the product, not the plumbing.” (claude.com) ### Why wasn’t one agent enough? Because long jobs break in boring ways. Context fills up. sessions get messy. agents lose track of what(thenewstack.io) as much as raw model quality, and that multi-agent setups help especially when work needs breadth, parallel exploration, or separate evaluator roles. Its own Research product already uses an orchestrator-worker pattern for that reason. (anthropic.com) ### What changed on May 6? Three things. First, “dreaming” adds a scheduled memory proces(claude.com)ites better distilled memories for future runs. Second, “outcomes” lets developers specify what good looks like, then uses a separate grader agent to judge whether the result actually met the bar. Third, multi-agent orchestration lets one agent spin up and direct other agents in parallel. (thenewstack.io) ### What does “dreaming” really mean? Not consciousness — just memory maintenance. (anthropic.com)s stay trapped inside messy session logs. Dreaming turns recent runs into a cleanup pass. The system revisits what happened, extracts patterns, and stores more useful memories, with controls for either automatic updates or human review before changes are written. (thenewstack.io) ### Why add outcomes and a grader? Because many real tasks are not just “produce text.” Th(thenewstack.io)t miss edge cases.” Anthropic’s answer is to make success criteria explicit, then have a separate agent grade the work against that rubric. That setup matters for detail-heavy jobs and also for fuzzier work like brand-consistent writing. In Anthropic’s testing, this improved task success by up to 10 points versus a standard prompting loop. (claude.com) ### Why is multi-agent orche(thenewstack.io)line of thought at a time. An orchestrator can split a problem into subproblems, hand them to separate workers with their own context windows, then combine the results. Think less “better chatbot” and more “project manager with specialists.” Anthropic has already used that pattern internally for research, and now it’s pushing the same idea toward developers as a product feature. (anthropic.com) ### What’s the catch? More autonomy means more syste(claude.com) harness assumptions go stale as models improve, which means developers still need good tools, clear rubrics, and strong permissions. Managed Agents removes a lot of infrastructure work, but it does not remove the need to decide what the agent should be allowed to do and how success gets checked. (anthropic.com) ### Bottom line? Anthropic is not just shipping a smarter model here. It’s productizing the operating system around agents — me(anthropic.com)nterprise AI shifts from prompt engineering toward managing teams of agents that can plan, delegate, and improve between runs. (claude.com)