Anthropic has quietly shipped a major upgrade to its Managed Agents platform, introducing a new agent type that allows Claude to write its own orchestration logic. The multiagent_20261001 type, documented as of October 10, 2026, enables agents to not only delegate to subagents but also generate dynamic workflowsβ€”programs that run multiple agents in parallel to handle complex, multi-piece tasks like reviewing hundreds of documents.

The Shift from Delegation to Orchestration

Previously, multiagent setups relied heavily on static delegation where the primary agent would hand off tasks to predefined subagents. The new system introduces "dynamic workflows," where the agent writes a program to coordinate other agents in the background. This allows for parallelization and specialization without blocking the main session thread. The agent can determine when to use a workflow based on instructions in its system prompt, effectively becoming the architect of its own computational strategy.

Subagents, Inline Agents, and Advisors

The documentation details three distinct coordination mechanisms: subagents, dynamic workflows, and an "advisor" model. Subagents remain persistent threads that the primary agent can message back and forth with, retaining their conversation history. Dynamic workflows, by contrast, create ephemeral threads that are archived by the server upon completion. Additionally, agents can now consult an "advisor"β€”a more capable model used for strategic guidance mid-turn, such as planning an approach or reviewing work before finalizing.

Configuration and Constraints

Developers must explicitly enable these features in the agent's definition block. By default, both subagents and workflows are enabled for the new type. The system supports both predefined agents (saved configurations) and inline agents (defined on the fly). However, strict validation rules apply: agents cannot be nested more than one level deep, and any agent listed in a workflow or subagent configuration must match the primary agent's inference geography. A mismatch results in a 400 validation error, ensuring consistent routing and compliance.

Key Takeaways

  • New multiagent_20261001 type enables agents to write their own workflow programs.
  • Dynamic workflows allow parallel execution of multiple agents in the background.
  • An "advisor" model can be consulted for strategic guidance during a turn.
  • Inline agents use the primary agent's model and tools, while predefined agents use their own.
  • One-level nesting limit and inference geography matching are strictly enforced.

The Bottom Line

This is a massive leap for autonomous coding. By letting Claude write its own orchestration code, Anthropic is moving beyond simple tool-calling into true self-directed engineering, though the strict geography and nesting constraints will test enterprise deployments.