Claude Managed Agents adds dynamic workflows for parallel jobs

Anthropic added dynamic workflows to Claude Managed Agents in beta on October 9, 2026. The release note describes an agent writing a program that runs other agents in phases, then combines their results. Anthropic executes that program in the background as a workflow run.

For developers, the useful question is how much coordination their application can delegate while keeping the result auditable. A document review service needs to account for every file it accepts, including the ones that fail. A finished report alone cannot provide that accounting.

Who can use the beta

Managed Agents supplies the agent harness and execution infrastructure. Developers configure the agent and its environment, start a session and exchange events with it. The current overview says access is enabled by default for Claude API accounts. Requests need an API key and the Managed Agents beta header, which the SDK sets automatically. Cloud platform implementations can differ, so teams should verify their chosen deployment separately.

How a workflow changes delegation

With ordinary subagents, the main agent decides the next step and can send further messages. With dynamic workflows, generated code passes work and results between agents while the main conversation remains available. The orchestration guide says workflow threads cannot receive those further messages and are archived by the end of the run. That makes the choice of coordination style consequential.

Consider a hypothetical product documentation review. An application could request a first phase that checks individual manuals, followed by a phase that reconciles inconsistent instructions. It should require a list of covered documents, unresolved conflicts and missing inputs. This is an illustrative use of the new capability. The release provides no independent measurement of how reliably this particular review would perform.

Migration deserves an explicit configuration review

Existing integrations should inspect the full multiagent configuration before changing to multiagent_20261001. Anthropic says the migration replaces that entire block. Preserve the intended agent list and advisor settings, and explicitly disable capabilities you do not want. Workflows and inline agents otherwise default to enabled. Existing sessions retain their configuration, so testing the change requires a new session. The older coordinator type remains accepted, but cannot start workflow runs.

A practical review should compare the saved configuration against the intended one, then exercise a small task with a known answer before expanding the workload. That creates a baseline for deciding whether the added coordination helps this application.

Track coverage as well as completion

A completed result means the workflow program finished. Individual work may still have failed. Anthropic’s run reference says to inspect thread events. Track every created run through workflow_run.status_ended, then wait for a session idle event with end_turn that your own interruption did not cause. An idle session alone is insufficient.

The documented limits are 64 working threads per run, a value Anthropic explicitly does not guarantee, and 1,000 agent starts during a run. The default lifetime is 24 hours. A session allows 10 open runs by default, including paused ones. Retries can produce additional thread records.

A normal interrupt ends the primary turn without ending its runs. They may pause or continue. Applications must account for that behavior in their stop controls. Workflow events arrive through the session stream. Answer tool requests as they arrive, even while the session is running. Replacing the session agent is blocked while any run remains open. Tools also need to tolerate repeated calls because work can be retried.

For the hypothetical documentation service, maintain an input inventory outside the generated summary. Compare it with the returned coverage list before delivering the report. A missing manual should remain visible as a gap. This acceptance rule is a design recommendation and would need testing against the actual application.

Set spend controls before the first run

The budget documentation says a session budget must be attached at creation. It cannot be added later to an uncapped session. All threads share the same allowance. Enforcement stops new model requests after the threshold, while requests already admitted finish. A parallel run can therefore exceed the threshold by one model request per working thread. Teams should leave room for that margin instead of treating the setting as an exact maximum invoice.

The comparison uses public list cost, including tokens, searches and runtime. Negotiated billing can differ. Reaching the allowance pauses work. Raising or removing the budget resumes work paused by the budget, subject to separate interruptions. Removing it is irreversible for that session, so adjusting the amount preserves a control that removing it gives up.

Anthropic’s published pricing charges model tokens plus $0.08 per running session hour. Idle time does not accrue that runtime fee. Web search costs $10 per 1,000 searches, alongside token usage. Managed Agents does not receive the Batch API discount. These are separate cost components, and the chosen model and amount of work determine the token bill.

For a useful pilot, record cost for a completed, reviewed result. Include failed inputs and repeated work in the accounting. Comparing only the first response time would miss the effort needed to make the output usable. This article is based on Anthropic’s documentation and does not report a hands on evaluation.

Review tool permissions and retained data

The permission system covers tools executed by Anthropic, including agent and MCP toolsets. Policies can permit execution, require approval or let the server assess each call. The defaults differ. Agent tools default to automatic permission, while MCP tools default to asking. Custom tools executed by an application remain that application’s responsibility. A team enabling parallel work should review the actions available to each worker before connecting systems that can change external records.

Data handling is another deployment constraint. Managed Agents keeps conversation history, sandbox state and outputs. The overview says it is currently ineligible for Zero Data Retention and HIPAA Business Associate Agreement coverage. Teams whose data requirements depend on either should resolve that mismatch before uploading a document collection.

The October release belongs to Managed Agents

Anthropic introduced dynamic workflows in Claude Code in an article dated May 28. It had also announced multiagent orchestration for Managed Agents on May 19. Earlier Claude Code availability statements and performance anecdotes should not be treated as evidence for this Managed Agents addition.

The strongest initial use case is a bounded collection of work with explicit coverage and review criteria. Start with documents you can check, decide which actions need approval and measure the result after review. The new orchestration capability makes those operating decisions more important as the number of workers grows.