At the opening of Code with Claude SF on May 6 in San Francisco, Anthropic spent little time on framing and went straight to three new Managed Agents features: Dreaming, now in research preview; Outcomes, now in public beta; and Multiagent Orchestration, also in public beta. Dreaming, the feature that lets Claude learn between tasks, had already drawn attention. This release puts the spotlight on the other major piece: multi-agent orchestration.
In plain terms, Claude Agent can now split itself into a working team. A lead agent receives a complex assignment, breaks it into parts, and hands those parts to specialized subagents that run in parallel. Each subagent can have its own model, prompt and tool set.
That may sound like outsourcing software work to a set of AI contractors, but several numbers make the launch worth a closer look.
Up to 20 agents, 25 threads and one shared file system
Anthropic's limit is specific: up to 20 different agents and up to 25 concurrent threads. This is not a small demo limit. It is the kind of capacity target meant for enterprise production workflows.
The architecture is specific as well. Subagents share the same file system, events are persisted, and the lead agent can return to a previous subagent mid-task and ask it to continue. It brings a familiar concurrency model from traditional multithreaded programming into the world of AI agents.
Anthropic's own example is an incident investigation. A lead agent receives the task, then subagents inspect deployment history, error logs, monitoring metrics and support tickets in parallel. Their findings are then merged back into the lead agent's context.
In a workflow where one agent would previously have to read hundreds of logs in sequence, 20 specialized agents working at once could, at least in theory, cut the elapsed time by an order of magnitude.
Netflix's platform team is already using it
Anthropic also gave unusually concrete customer examples instead of leaning on vague references to unnamed enterprise users.
Netflix's platform team has used multiagent orchestration to build an analysis agent that processes logs from hundreds of builds and looks for recurring classes of problems across applications. It is a classic case of work that is high-volume but still requires close inspection: too subtle for a conventional script, too slow for a single agent, and therefore well suited to a multi-agent setup.
Harvey, the legal AI company, used Dreaming to remember tool-calling patterns and file-type pitfalls from document workflows, improving completion rates by roughly sixfold. That is not a vague claim of meaningful improvement; it is a specific number for legal document coordination.
Spiral uses a more cost-aware mix: Haiku handles the first request because it is cheaper and faster, Opus writes the finished draft because it is stronger, and Outcomes enforces editing standards between the two. The result is a workflow that can tune cost and quality together.
Wisedocs reported a 50% speed improvement in document quality review while using outcome scoring to keep team standards consistent.
The common thread is that these customers treat multi-agent collaboration as an engineering problem, not a research display. Enterprises will care less about the phrase multiagent orchestration than about whether the system runs reliably, controls cost and reduces manual follow-up after launch.
Outcomes may be the hidden lever
The multi-agent story is easy to tell. But in Anthropic's release, Outcomes may be the more important tool for attacking the old problem of AI systems doing unreliable work.
The design is direct: developers write a rubric for success, and a separate grader evaluates the agent's output in an isolated context. The grader does not see the agent's reasoning process; it only sees the result. If the output fails, feedback is sent back and the agent tries again. The default is three retries, with a maximum of 20.
Anthropic's own test results show measurable gains:
| Task type | Performance lift |
|---|---|
| docx file generation | +8.4% |
| pptx file generation | +10.1% |
A 10 percentage point gain is meaningful for workflows already running in production, especially when the starting baseline is not low.
The mechanism matters even more than the numbers. Outcomes separates evaluation from generation. In RLHF and self-critique patterns, the same model often acts as both athlete and referee, which can lead it to convince itself that an imperfect answer is good enough. Outcomes replaces that loop with another agent that does not know how the first one reasoned, cutting off that form of self-justification.
Why this matters
Seen one week at a time, Anthropic's rollout can look like a stack of feature launches. Taken together, the pattern is clearer:
- Dreaming is about accumulating memory.
- Outcomes is about quality assurance.
- Multiagent Orchestration is about splitting complex work into parts.
Together, they point to a product shape in which an enterprise agent can run real work independently without constant human supervision. That is a different class of product from the one-question, one-answer chat interface.
At Code with Claude, Anthropic repeatedly used the phrase "agentic by default." From product decisions to pricing, including the agent subscription models now watched closely on Wall Street, the company's strategy is moving in that direction.
Whether this setup holds up in large enterprise deployments remains to be seen. Beyond Harvey's sixfold improvement, Wisedocs' 50% speed gain and Netflix's hundreds of builds, there are still hard operational questions around recovery from errors, cost controls and security boundaries.
But one conclusion is already hard to avoid: the era of the single chat box is ending.
Sources: New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration (Claude Blog); Anthropic updates Claude Managed Agents with three new features (9to5Mac); CocoLoop; Claude's new "Dreaming" feature is designed to let AI agents learn from their mistakes (The Decoder); Anthropic just taught Claude to dream between tasks, and it makes agents meaningfully smarter (Digital Trends)