Anthropic just enabled dynamic workflows for its Claude Managed Agents, a move that pushes agentic systems beyond simple tool-use into autonomous, multi-step execution. An agent can now write and run its own programs in the background to orchestrate up to a thousand sub-agents for complex tasks.
This isn't just another API feature; it's a shift in how we can architect and deploy systems that handle work at scale, like processing hundreds of documents or auditing large codebases.
what just shipped
On October 9, in its platform release notes, Anthropic detailed the new beta feature. Instead of a developer defining a rigid chain of tasks, a managed agent can now dynamically write a workflow—a program that runs other agents in phases and combines their results. This entire workflow then executes on Anthropic's servers as a background run, allowing the primary user-facing session to continue without being blocked.
The scale is significant. A single workflow run can orchestrate up to 1,000 agents over its lifetime, with as many as 64 agents working concurrently. In an internal test, Anthropic reported that a single agent found between 14 and 27 of 70 bugs planted in a large codebase, whereas a dynamic workflow consistently found 66.
why this is a step-change for builders
This model moves past simple delegation. The key distinction is that the lead agent writes code to orchestrate the other agents, passing context and results between them without its own direct involvement in every step. The orchestration logic itself is a program, not just another model turn.
More importantly, the agent decides for itself when to initiate a run based on the user's request and its system prompt. This autonomy is a fundamental change. Your prompt is no longer just a request for an answer; it's a potential trigger for a complex, multi-agent process that runs independently in the cloud.
For anyone building production agents, this pattern solves the problem of long-running, complex tasks that would otherwise time out or require a complex state-management system on the client side.
how to use it (and the gotchas)
The feature is in beta and must be explicitly enabled. It's accessible via the Claude Developer Platform by using a specific beta header (managed-agents-2026-04-01). To configure it, you enable workflows within your agent's multiagent settings.
{
"multiagent": {
"type": "multiagent_20261001",
"workflows": {
"type": "enabled"
}
}
}
Once enabled, you guide the agent on when to use workflows via its system prompt. You can then monitor the progress through workflow_run events in your session's event stream.
The primary gotcha is cost and control. While there's no fee for the workflow run itself, you are billed for the tokens used by every agent in the run, plus a session runtime fee. Since the agent decides when to start a run, setting a budget is the main brake on a process that could otherwise consume significant resources.
the takeaway
Anthropic's dynamic workflows are a strong signal of where the industry is heading. We are moving from single-shot, human-in-the-loop tools toward persistent, autonomous agents that can manage and execute complex, long-running jobs.
For developers, this means thinking less about managing individual API calls and more about designing systems of agents and the guardrails they operate within. It's a more powerful but also more abstract way of building, and it's here now.
Top comments (0)