Claude Managed Agents adds dreaming research preview, webhooks, and multiagent orchestration

• Anthropic added four concrete features to Claude Managed Agents, according to ClaudeDevs' launch thread: multi-agent orchestration, an outcomes loop for rubric-driven self-improvement, dreaming as a research preview, and webhooks.
• In ClaudeDevs' multi-agent post, Anthropic says specialized sub-agents get their own context windows while sharing one container and filesystem, a setup aimed at longer coordinated runs.
• ClaudeDevs' outcomes post says a dedicated grader sub-agent can score work against a rubric and feed that result into the next iteration.
• With ClaudeDevs' webhook post, agents can push status updates to workspace webhooks instead of relying on an open SSE stream or polling.
• ClaudeDevs' dreaming post describes dreaming as a background process that reviews many sessions, edits memories, adds skills, and removes stale ones.

Anthropic linked a multi-agent docs page, an outcomes docs page, a dreams docs page, and webhooks docs straight from the launch thread. It also pointed developers to a Claude Code skill for getting started, which makes this feel less like a vague research tease and more like a toolkit drop with one especially experimental piece.

Anthropic's clearest systems change is in ClaudeDevs' multi-agent post: sub-agents can delegate work to specialized peers, each with its own context window, while still coordinating through a shared container and filesystem.

That shared workspace is the useful detail. Separate context windows keep specialist runs from stepping on each other's token budgets, while the shared filesystem gives them a place to hand off artifacts, logs, and intermediate outputs. Anthropic's linked multi-agent docs suggest this is meant for agent teams, not just one model calling tools in a loop.

The outcomes feature is Anthropic's built-in critique pass. In ClaudeDevs' outcomes post, the company says a grader sub-agent evaluates completed work against a rubric, then returns feedback for another iteration.

That gives Managed Agents a native pattern for rubric-based revision:

• define the task-completion rubric
• run the worker agent
• hand the result to a grader sub-agent
• feed the grader's feedback into the next pass

Anthropic's outcomes documentation is the official source here, but the tweet already surfaces the core architecture: one sub-agent does the work, another judges it.

Dreaming is the research-preview headline. In ClaudeDevs' dreaming post, Anthropic says agents write memories during a session, then a background process reflects across many sessions to curate those memories by editing them, adding new skills, or removing stale ones.

Anthropic's wording matters because it goes beyond retrieval. The linked dreams documentation frames the system as memory curation over time, not just fetching old context on demand.

Aakash Gupta's reaction in aakashgupta's forgetting thread pushes on the part Anthropic only hints at: pruning. He argues that most agent memory systems keep appending to a vector store, while dreaming implies periodic editing and deletion so the memory corpus does not turn into a slowly noisier log.

Anthropic also shipped a plumbing upgrade that will matter to anyone running long jobs in production. According to ClaudeDevs' webhook post, agents can send updates to webhooks configured in a Claude workspace, which removes the need to hold open an SSE stream or poll for completion.

The linked webhooks docs make this the least glamorous feature in the bundle and maybe the most immediately deployable one. It gives Managed Agents a cleaner path into external automations, dashboards, and job runners.

The most interesting outside read came from aakashgupta's initial dreaming thread and his follow-up on forgetting, which both frame dreaming as offline consolidation plus pruning, not just another memory layer.

His argument breaks into three concrete claims:

• retrieval-only memory stores facts, but does not extract reusable patterns
• curation implies deleting or rewriting weak memories, not keeping every session forever
• bigger context windows do not solve the noise problem by themselves

Those claims are commentary, not Anthropic documentation. But they fit the one verb Anthropic chose in its dreaming description: curate.

Anthropic did not bury the entry points. In ClaudeDevs' getting-started post, it told developers to try these features through the /claude-api skill in Claude Code or the open source repo.

The linked GitHub skill is the practical bridge between the research-preview framing and actual use. The same launch sequence also landed during Anthropic's Code with Claude event, where ClaudeDevs' event schedule post put Boris Cherny and Jarred Sumner on the build-with-Claude-Code segment later that day.

That timing adds one more concrete signal: Anthropic presented Managed Agents as part of a live coding workflow story, not as a detached research demo.