2026-05-07
Anthropic Managed Agents adds fleet orchestration, Outcomes, and self-improving Dreaming
Anthropic ships fleet orchestration, Outcomes goal-spec, and Dreaming for Managed Agents. Dreaming lets agents self-improve across sessions without fine-tuning.
At the Code with Claude SF conference on May 6, Anthropic announced three new capabilities for Claude Managed Agents that materially change how long-horizon agentic workflows are built.
Multi-agent fleet orchestration
Managed Agents now supports spinning up coordinated fleets of sub-agents to parallelize subtasks. You define a high-level task; Claude decomposes it, spawns multiple agents to work on distinct subtasks simultaneously, and synthesizes their outputs. This is similar in concept to what you’d build manually with an orchestrator + worker pattern in the Agents SDK — but Anthropic’s implementation handles failure recovery, output reconciliation, and context sharing across the fleet automatically.
The practical effect: tasks that previously required sequential agent loops (because managing parallel state was too complex) can now be parallelized without custom orchestration code.
Outcomes: goal-driven iteration
Outcomes is a goal-specification layer. Instead of defining every intermediate step in a workflow, you describe what success looks like — a test suite that passes, a PR that addresses all reviewer comments, a docs page that accurately reflects the current API — and Claude iterates until the success criterion is met.
This shifts the developer’s role from workflow engineer to success-criterion author. The number of intermediate steps, retries, and tool calls are determined by Claude, not by your prompt structure.
Dreaming: session-to-session self-improvement
Dreaming is the most novel of the three. After a session ends, Claude reviews its own session history, identifies mistakes or gaps, and stores structured observations for use in the next session against the same repo or task type. This is model-initiated memory — the agent improves its performance on your specific codebase over time without developer-plumbed context injection, RAG pipelines, or fine-tuning.
The name references the neuroscience hypothesis that dreaming consolidates learning from the day’s experiences. In practice it means agents accumulate knowledge about your codebase’s idioms, failure modes, and conventions through repeated runs — rather than starting from scratch each session.
Practitioner note
Outcomes is the most immediately useful feature for teams with well-defined success criteria. If your agent workflow ends with “run the test suite and open a PR if green,” you now have a built-in goal spec rather than a bespoke retry loop. For Dreaming: the value compounds with volume — an agent that runs 15 times against the same repo will meaningfully outperform one that runs once. If you’re on Max plan (15 Routines/day), enabling Dreaming on a high-frequency routine is worth doing from day one. Fleet orchestration is most valuable for tasks that are naturally parallelizable (e.g., reviewing 20 files simultaneously rather than sequentially) — identify those in your workflow first.
Sources
- Code with Claude SF 2026 recap — Simon Willison ↗
- Code with Claude SF recap — Blake Crosley ↗
- Code with Claude conference page ↗