Anthropic shipped two new pieces to Claude Managed Agents at last week's Code with Claude 2026 conference: Multiagent Orchestration (lead agent + specialist sub-agents) and Outcomes (a rubric-graded iteration loop). Both are public beta now. For anyone building agent workflows past the "one prompt, one agent, one task" pattern โ and a lot of teams have been hitting that ceiling on complex investigations or multi-step content generation โ these are the orchestration primitives Anthropic was missing versus where LangChain, CrewAI, and AutoGen have been operating.
Multiagent Orchestration: a lead agent breaks complex tasks into pieces and delegates to specialist sub-agents, each with its own model, prompt, and tools. Sub-agents work in parallel on a shared filesystem and contribute back into the lead's context. Persistent event memory spans the whole fleet, with full tracing in Claude Console showing which agent did what, when, and why. The example Anthropic cited from customer Spiral is the right shape: Haiku as lead agent for cheap triage and request routing, Opus instances delegated for drafting โ model heterogeneity is the point, not single-model swarms. Outcomes adds a separate Claude instance as grader: you write a rubric describing what success looks like, the grader evaluates output in its own context window (isolated from the agent's reasoning trajectory), and when the grader pinpoints issues the agent iterates. Reported gains: up to 10 percentage points on hardest tasks vs a standard prompting loop, with specific numbers of +8.4% on docx generation and +10.1% on pptx. The grader-in-separate-context architecture is the genuinely new bit โ it isolates the success metric from the same model that produced the work, closer to LLM-as-judge harness territory than a chain-of-thought self-critique.
Multi-agent patterns have been in the open-source agent stack for over a year โ LangGraph, CrewAI, AutoGen, Microsoft's AutoGen Studio โ so Anthropic is late to ship a managed version. But "late and integrated" beats "early and stitch-together-yourself" for a lot of teams: persistent event memory + Console tracing + shared filesystem + first-party access to Claude models removes orchestration glue that previously sat in user-maintained Python or someone's leaky abstraction. Outcomes is the more architecturally interesting piece because it changes what an evaluation loop looks like inside production agent workflows. Standard prompting loops bake the grader into the same context as the agent, which means the agent's own reasoning trajectory steers what gets "graded as good" โ and you end up with self-consistency dressed up as quality control. Splitting the grader into its own context (same model family, different instance) gives you LLM-as-judge inside the agent's runtime, not as an offline eval. The 10-percentage-point gain claim is specific enough to test against your own workload before believing it, but the architecture matches what works in research literature.
Both features are in public beta โ no waitlist for Outcomes or Multiagent Orchestration. Dreaming (the separate memory-curation feature also announced) still requires request access. Console-visible from day one, so the operational tooling is real, not vapor. If you're running Claude agents and finding that "one big prompt with tool use" hits a ceiling on complex tasks, Multiagent Orchestration is where to start โ the Haiku-leads-Opus pattern from Spiral is a copyable shape. If you're generating structured output (docs, presentations, code) where quality matters more than throughput, Outcomes is where the percentage points live. Pricing wasn't disclosed in the announcement, so the cost-per-task math against a single-agent loop is the next thing to figure out before going to production. Worth running A/B against your current workflow before committing.