Claude Managed Agents: Dreaming, Outcomes & Orchestration

At its May 6 Code with Claude conference, Anthropic shipped three new Managed Agents API features that change how production agents are built: Outcomes, Multiagent Orchestration, and Dreaming. Two of them — Outcomes and Multiagent Orchestration — are in public beta right now with no waitlist. Dreaming is more restricted. If you are building production agents on Claude, here is what actually changed and what to do with it this week.

Outcomes: Let a Grader Agent Verify the Work

The hardest problem in agent development is not building the agent — it is knowing whether it actually completed the task correctly. Outcomes is Anthropic’s answer to that.

The model is simple: you write a markdown rubric describing what success looks like, upload it, and attach it to the agent session. When the agent produces output, a separate grader agent evaluates it against your rubric in its own context window. That isolation matters — the grader has no access to the agent’s reasoning chain, so it cannot be talked into accepting bad output. If the rubric is not satisfied, the grader sends targeted feedback back to the agent, which takes another pass. You set a maximum iteration count (five is a reasonable default).

# Upload your rubric file
rubric = client.beta.files.upload(file=Path("/tmp/rubric.md"))

# Attach an outcome to the session
outcome = {
    "type": "user.define_outcome",
    "description": "Build a DCF model for Costco in .xlsx",
    "rubric": {"type": "file", "file_id": rubric.id},
    "max_iterations": 5
}

The required beta header is anthropic-beta: managed-agents-2026-04-01. No waitlist, no separate access request. MindStudio ran a production test using Outcomes on PowerPoint generation and measured a 10.1 percent quality improvement. The pattern works.

Outcomes is the feature to implement this week. If your agents produce any structured deliverable — reports, spreadsheets, code, summaries — you now have a principled way to verify quality instead of spot-checking output manually.

Multiagent Orchestration: Stop Cramming Everything Into One Context

Long-running agent tasks hit a ceiling when the job is too complex for a single context window to handle well. Multiagent Orchestration lets you break the ceiling.

A lead agent decomposes the task and delegates pieces to specialist subagents. Each specialist runs with its own model, system prompt, and tool set — isolated context, purpose-built configuration. They execute in parallel on a shared filesystem. The lead agent can check in with any of them mid-workflow because events are persistent and every agent maintains its own history of what it has done.

The observability story is unusually good here. The Claude Console shows a full trace of the entire multiagent session: which agent handled which step, in what order, and the reasoning behind each delegation decision. If something goes wrong, you can see exactly where.

A practical pattern from the Anthropic docs: a lead agent runs an incident investigation while subagents simultaneously query deploy history, parse error logs, check metrics dashboards, and pull support tickets — all in parallel. Previously building this required maintaining your own orchestration layer with routing logic, state management, and error handling across agents. Now you configure it in the session YAML and let the API handle coordination.

Like Outcomes, Multiagent Orchestration is in public beta and available via the managed-agents-2026-04-01 beta header. No separate access request required.

Dreaming: Agents That Improve Between Sessions

Dreaming is the most ambitious of the three features and the most restricted. It is in research preview — you need to request access at claude.com/form/claude-managed-agents before you can use it.

A scheduled background process reviews your agent’s past sessions and extracts behavioral patterns: recurring mistakes, efficient workflows the agent has converged on, preferences shared across a team. It then curates updates to the memory store so your agent gets better at your specific workflows without retraining or manual prompting. Harvey, a legal AI platform, reported that task completion rates increased approximately 6x after implementing Dreaming to consolidate agent memory across legal research sessions.

dream = client.beta.dreams.create(
    memory_store_id="ms_...",
    session_ids=["session_1", "session_2"],
    model="claude-sonnet-4-6"
)
# Required header: anthropic-beta: managed-agents-2026-04-01,dreaming-2026-04-21

You decide the level of control: Dreaming can apply memory updates automatically, or it can queue proposed changes for your review before anything lands. Supported models during research preview are claude-opus-4-7 and claude-sonnet-4-6. This is not ready for everyone today — but the waitlist is real and the concept is sound. Apply now so you are not blocked when you need it.

The Rate Limit Context That Makes This Practical

Also announced at Code with Claude: Anthropic’s deal with SpaceX allocates the entire Colossus supercluster (220,000+ GPUs) to Claude. Rate limits for subscription plans doubled; peak-hour throttling is gone for Pro and Max accounts; API rate limits are up to 17x higher for some tiers. Multiagent workloads previously hit rate limits fast enough to make parallel orchestration impractical. According to Anthropic’s announcement, that constraint has now eased substantially.

Webhooks Close the Async Loop

Webhooks — also new, also in public beta — send HTTP callbacks whenever a significant agent state change occurs: session starts, agent idles waiting for input, outcome evaluation completes, a multiagent thread terminates, a stored credential refresh fails. No more polling the API to check on long-running sessions.

Combined with Claude Code Routines (released in research preview April 14), you can chain fully automated pipelines: a deploy event fires a webhook that triggers an agent routine, the routine runs smoke tests, and a completion webhook posts results to Slack. No human in the loop until something fails.

What to Do This Week

• Enable Outcomes today — Add managed-agents-2026-04-01 to your beta header, write a rubric for your highest-stakes agent output, and attach it. Full docs at platform.claude.com.
• Evaluate Multiagent Orchestration — If you have any agent workflow running tasks sequentially that could run in parallel, read the multiagent session docs and start with the incident investigation pattern as a template.
• Request Dreaming access now — Even if you do not need it yet, the waitlist is real. Apply at claude.com/form/claude-managed-agents before demand builds further.