Dynamic workflows are a beta feature of Claude Managed Agents, Anthropic's hosted agent harness on the Claude API. The agent writes a workflow — a program that runs many agents in phases and combines their results — and Anthropic's servers execute it in the background as a workflow run. Per the official docs, one run can start up to 1,000 agents over its life with 64 working at once, runs last 24 hours by default, and there is no separate price: you pay normal token rates plus $0.08 per session-hour of runtime.
Last updated October 12, 2026 · Sources checked October 12, 2026
Key facts
| What it is | A Claude Managed Agents feature: the agent writes a program that fans work out to many agents and combines results |
|---|---|
| Status | Beta (header managed-agents-2026-04-01) |
| How to turn it on | multiagent.type: "multiagent_20261001" — workflows and subagents are on by default with this type |
| Agents per run | Up to 1,000 started over the run's life (then thread_limit_error) |
| Concurrency | 64 threads working at once per run (not guaranteed, may change) |
| Run lifetime | 24 hours by default, or what the agent sets |
| Open runs per session | 10 by default |
| Predefined agents | Up to 20 per list |
| Price | No run fee — tokens at model rates + $0.08 per session-hour while running |
| Who can start a run | Only the agent on the session's primary thread (runs don't nest) |
What dynamic workflows are
Claude Managed Agents gives you a pre-built agent harness and sandboxed infrastructure instead of writing your own agent loop. With the multiagent_20261001 agent type, the agent can hand work to other agents in two ways:
- Subagents — the agent delegates tasks itself and reads each report. A session can have at most 25 child threads at a time.
- Dynamic workflows — the agent writes a workflow program that runs many agents in phases in the background. Anthropic suggests it for “audits, migrations, deep research, and cross-checking”.
You decide which mechanisms the agent may use; the agent decides when to use them. Anthropic recommends telling it in the system prompt when a workflow run is appropriate, and you can restrict it to agents you list (up to 20).
Limits from the docs
The Workflow runs page lists these limits:
| Limit | Value | At the limit |
|---|---|---|
| Threads working at once in one run | 64 | No new threads until one finishes; not guaranteed |
| Agents a workflow starts over the run's life | 1,000 | Run ends with thread_limit_error |
| Run lifetime | 24 h default | Run ends with timeout_error |
| Runs open at once in a session | 10 default | Start refused (max_workflow_runs_error) |
A run's threads don't count against the session's 25-child-thread limit, and the server archives them itself. Runs can't nest: only the primary-thread agent can start one.
How much dynamic workflows cost
A workflow run has no price of its own. Per Anthropic's pricing page, Managed Agents bills tokens at normal model rates (prompt-caching multipliers apply; the Batch API discount does not) plus session runtime at $0.08 per session-hour, metered only while the session is running. Every agent in a run uses tokens, so a 1,000-agent run can spend a lot quickly.
Cap it with a session budget. When a session hits its budget every open run pauses and the session goes idle with budget_reached; raising or removing the budget resumes them. Each working thread finishes its in-flight request, so you can overshoot by up to one request per thread.
Dynamic workflows vs Claude Code
They are different products. Claude Code is Anthropic's coding agent for your terminal, IDE, desktop and web, included with Claude Pro and Max subscriptions. Claude Managed Agents is an API platform for developers building their own agents on Anthropic-hosted (or self-hosted) sandboxes, billed per token and runtime. Dynamic workflows live in Managed Agents; to use them you call the API (or the SDK / ant CLI), not the Claude Code app.
If you just want parallel coding help on your own repo, Claude Code, Codex cloud tasks or Devin are simpler. Dynamic workflows make sense when you are building a product or pipeline that has to fan out over hundreds of files, documents or checks.
How to enable it
Create an agent with the beta header and a multiagent block (adapted from the docs):
curl https://api.anthropic.com/v1/agents \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "anthropic-beta: managed-agents-2026-04-01" \
-H "content-type: application/json" \
-d '{"name": "auditor", "model": "claude-opus-5-5",
"tools": [{"type": "agent_toolset_20260401"}],
"multiagent": {"type": "multiagent_20261001",
"subagents": {"type": "disabled"}}}'Workflows are on by default for this type; the subagents: disabled line keeps only workflows, as the docs describe. Follow progress on the session event stream: workflow_run.created, workflow_run.status_running, workflow_run.phase_started / phase_ended, workflow_run.status_idle (e.g. at the budget), workflow_run.error and workflow_run.status_ended. Check the multiagent orchestration page for the exact model IDs and SDK calls.
About the “66 of 70 bugs” claim
Launch coverage (for example AI Understanding, Oct 11, 2026) and widely shared X posts say Anthropic planted 70 bugs in a 116,000-line codebase: a single agent found 14, 15 and 27 across three runs while a dynamic workflow found 66 each time. We could not find those numbers in the API docs we read, and the coverage notes that time, token cost and false-positive rates were not disclosed. Treat it as a vendor demo, not a benchmark.
Sources
- Claude Managed Agents overview — Claude API docs
- Multiagent orchestration — Claude API docs
- Workflow runs (states, events, budgets and limits) — Claude API docs
- Session threads (25 child-thread limit) — Claude API docs
- Claude Managed Agents pricing — Claude API docs
- AI Understanding — “Anthropic launches Claude Managed Agents dynamic workflow beta” (Oct 11, 2026, secondary)