Claude Managed Agents Introduces Dynamic Workflows in Public Beta
Anthropic has launched dynamic workflows for Claude Managed Agents in public beta, allowing lead agents to orchestrate multi-agent tasks.
Claude Managed Agents dynamic workflows allow lead agents to orchestrate large-scale, parallelized tasks.
- Anthropic has released dynamic workflows for its Managed Agents system in public beta.
- The feature enables a lead agent to generate and manage complex multi-agent plans, capable of scaling up to 1,000 agents per run.
- Dynamic workflows allow for task segmentation, where agents work in phases to process complex workloads beyond the capacity of single sessions.
Overview of Dynamic Workflows
Anthropic has officially introduced dynamic workflows for Claude Managed Agents, now available in public beta. This update provides a new architectural approach to multi-agent orchestration, designed to handle tasks that exceed the scope or token capacity of a standard, single-thread AI interaction. By enabling a lead agent to construct and oversee a programmatic workflow, developers can automate complex, multi-stage operations that require specialized agent intervention.
How the Orchestration Functions
At the heart of dynamic workflows is the ability for a lead agent to function as a planner. When assigned a task, the agent defines a series of phases, each requiring different operations or agent capabilities. The system then executes these instructions in the background, spawning dedicated session threads as needed to perform specific parts of the work. This "fan-out" capability allows the system to engage multiple agents simultaneously, processing information across different threads before synthesizing the results for the primary agent.
The orchestration logic is handled by the server, which manages the lifecycle of these secondary agents. This process enables the primary agent to delegate high-volume tasks—such as reviewing hundreds of documents or executing exhaustive bug-hunting routines—without being constrained by the limits of a single conversation. Once a workflow run concludes, the results are returned to the primary session thread, allowing the lead agent to review the outcome and determine the next logical step in the process.
Technical Implementation and Control
To access this feature, developers must configure their agent with the multiagent_20261001 type and enable the workflows setting within the agent's multiagent block. The system supports substantial scale, with the capability to orchestrate up to 1,000 agents per individual run. Because these workflows can be resource-intensive, Anthropic recommends that users begin with scoped tasks to monitor token usage and performance before incrementally increasing complexity.
The platform provides granular control over the workflow execution, including:
- Phased Execution: Workflows can be divided into distinct stages, such as data retrieval, reconciliation, or analysis.
- State Management: Run states (running, idle, or ended) are tracked, with provisions for budget constraints and lifetime limits (defaulting to 24 hours).
- Failure Handling: The lead agent can be instructed on how to handle specific agent failures, either by retrying, adapting the plan, or terminating the run.
Use Case: Automated Bug Hunting
Demonstrations provided by Anthropic highlight the utility of this orchestration in software engineering. In a test involving a 116,000-line codebase, a single agent struggled to identify bugs consistently. By contrast, a dynamic workflow, which divided the codebase into manageable segments and utilized multiple agents to analyze them in parallel, identified a significantly higher percentage of bugs across repeated trials. This underscores the primary utility of the feature: enabling AI systems to decompose large-scale problems into parallelizable, manageable chunks.
Developers interested in exploring these capabilities can start by utilizing the /claude-api managed-agents-onboard bug-hunter command within the Claude Code CLI, or by visiting the official documentation for further configuration details.
Enjoyed this?
Get more posts like this delivered to your inbox.
