Anthropic Claude Managed Agents Can Write Dynamic Workflows That Run Up to 1000 Agents in Phases, in Beta
Anthropic put dynamic workflows in beta for Claude Managed Agents on Oct 9, 2026: the agent can write a program that runs many agents in phases, up to documented limits of about 1000 agents, 64 concurrent threads, and a 24-hour default lifetime.
Anthropic put dynamic workflows into beta for Claude Managed Agents on October 9, 2026. The feature lets an agent write a workflow program that fans work across many agents in phases, then combines their results while the main session thread keeps talking to you.
According to the Claude Platform release notes, you turn it on under the existing managed-agents-2026-04-01 beta header by setting the agent’s multiagent field to {"type": "multiagent_20261001", "workflows": {"type": "enabled"}}. With that type, workflows and subagents are both on by default.
The pitch is clear for large, pieceable jobs such as reviewing hundreds of documents. The catch is also clear: the agent decides when a run starts, permission policies do not gate that start, and every agent in a run burns tokens at list rates plus session runtime.
What dynamic workflows actually do
Anthropic’s multiagent orchestration guide draws a hard line between three approaches.
Subagents are still there: the lead agent delegates, can follow up, and stays one level deep with at most 25 child threads at a time (idle ones included). Advisor mode lets the primary thread consult a stronger model mid-turn while it keeps doing the work itself.
Dynamic workflows are different. Claude writes a program that orchestrates agents without Claude sitting in the middle of every handoff. Context and results move in code from one agent to another. The server runs that program in the background as a workflow run. The main thread can keep chatting with the user and check progress through workflow_run.* events.
A run has layers: the workflow run itself, named phases such as “Read the contracts,” and agent threads the server creates as the program needs them. The program can fan out agents in parallel, pass one agent’s output into the next, repeat steps until a review passes, and decide what to do when a worker fails.
You do not send a separate API call to start a run. You describe the work in a user.message. The agent decides whether and when to start. Anthropic says the practical lever is the system prompt: tell the agent which jobs deserve a run and which small jobs it should handle alone.
Confirmed limits versus soft edges
The workflow runs documentation lists hard ceilings and a few numbers it will not promise.
| Limit | Documented value | What happens at the edge |
|---|---|---|
| Threads working at once in one run | 64 | No new threads until one finishes. Anthropic says the API does not guarantee this number. |
| Agents a workflow starts over a run’s life | 1,000 | Run ends with thread_limit_error. Retries can mean more than 1,000 threads. |
| Run lifetime | 24 hours by default (agent can set shorter) | Ends with timeout_error. Paused time still counts. No event reports the lifetime the agent picked. |
| Runs open at once per session | 10 by default (idle included) | Start refused; you get max_workflow_runs_error. |
| Subagent child threads (separate path) | 25 at a time | Advisor threads and workflow-run threads do not count against this. |
Other server limits exist and are deliberately not fully listed. A run that breaks an unlisted rule can end in program_error or unknown_error. Archiving a session while a run is open might return 400 with workflow_run_open, or it might succeed and stop open runs.
There is still no list-runs endpoint. To price one run, you list session threads, keep those with the run’s workflow_run_id, and sum token usage. Archived threads stay in that list.
Confirmed versus unconfirmed
| Claim | Status | Source |
|---|---|---|
| Dynamic workflows in beta as of Oct 9, 2026 | Confirmed | Claude Platform release notes |
Enable with multiagent_20261001 and workflows.enabled |
Confirmed | Release notes and orchestration guide |
| Up to ~1,000 agents over a run life; 64 concurrent; 24h default lifetime | Confirmed (64 not guaranteed) | Workflow runs docs |
| Permission policies apply to tools, not to starting a run | Confirmed | Orchestration guide |
| Session runtime $0.08 USD per hour while status is running | Confirmed | Claude pricing |
| Exact dollar cost of a 1,000-agent production job | Unconfirmed | Depends on models, tokens, retries, and runtime; no published flat fee |
| GA date or removal of the beta header | Unconfirmed | Not stated in the Oct 9 notes |
How you turn it on, and how you brake it
Create or update an agent with multiagent.type set to multiagent_20261001. Workflows stay enabled unless you set workflows to {"type": "disabled"}. You can list up to 20 predefined agents a workflow may use, disable inline agents so only your list is allowed, or leave inline agents on so the workflow invents workers with the session agent’s model.
Inline agents use the session agent’s model. If you want cheaper workers for fan-out, create Haiku agents first and put them in workflows.predefined_agents. That matters for cost: Claude Opus 5.5 is $4 / $20 USD per million input and output tokens, while Claude Haiku 5.5 is $0.10 / $0.50 for prompts up to 100,000 tokens.
Session budgets are the main brake. Set a budget when you create the session. You cannot add one later. When spend hits the cap, open runs pause and the session goes idle with budget_reached. Each working thread still finishes the model request it already started, so a run can overshoot by up to one request per concurrent thread. Raise or remove the budget to resume, unless an interrupt also paused the run.
Permission policies still evaluate tool and MCP calls inside a run, including the auto policy Anthropic added earlier. Starting the fleet itself is not behind that gate. If that split makes you uneasy, disable workflows on agents that should never spawn runs.
Pricing in USD
Managed Agents bill two ways: standard model token rates for every thread, and session runtime at $0.08 USD per session-hour while status is running. Idle, rescheduling, and terminated time do not accrue runtime. Prompt caching multipliers apply. Web search inside a session still costs $10 per 1,000 searches. US-pinned inference adds a 1.1x token multiplier. There is no Batch API discount on Managed Agents sessions.
Anthropic’s worked example for a one-hour Claude Opus 5 session with 50,000 input and 15,000 output tokens totals $0.705 USD before caching. That is a single-agent sketch, not a 64-thread fan-out. A document review that spins dozens of workers will be dominated by tokens, not the $0.08 runtime line.
For Indian teams and other rupee-costed budgets, quote USD list rates first, then convert. The practical control is still a hard session budget plus Haiku workers for parallel read steps.
What it means for Indian developers
This is an API Platform feature, not a Claude.ai chat toggle. Indian startups already wiring Managed Agents for contract review, KYC packet triage, or codebase audits can cut custom orchestrators for the happy path: describe the job, let Claude write the fan-out program, and watch phases on the event stream.
Cost discipline matters more here than in a single-thread Claude Code session. A careless prompt that says “review every PDF in the folder” can launch a run that chews through Haiku or Opus tokens until the 1,000-agent or budget limit stops it. Put spend caps on every production session. Prefer Haiku 5.5 workers for bulk reads and keep Opus for reconcile and final answer steps.
Compliance teams should read the permission split carefully. Tool policies and allowed_hosts networking still matter after our coverage of Claude agent eval misuse, but they do not stop a run from starting. Pair that with clear system-prompt rules for when runs are allowed.
If you already use Claude Projects for Claude Code or product-facing agents, treat this as a different surface: Managed Agents are the harness with sandboxes, vaults, and now background workflow programs. See our note on the Claude Projects redesign for Claude Code for the product-side parallel-thread story, which is related in spirit but not the same API.
How it sits next to other agent news
Google’s Gemini Agent remains in private preview with Workspace and Microsoft 365 planning, and can route jobs to Claude in that product story. Anthropic’s bet here is lower-level and developer-facing: you own the agent definition, the sandbox, and the event stream. For cheaper parallel workers on Managed Agents, the same week’s Claude Haiku 5.5 pricing cut is the model most teams will list under workflows.predefined_agents.
Secondary reporting from Mixed News on October 11 matches the primary docs on the 1,000-agent ceiling, the permission-policy gap, and the budget overshoot rule. Stick to Anthropic’s pages for implementation details.
FAQ
When did Anthropic announce dynamic workflows?
October 9, 2026, in the Claude Platform release notes, as a beta under the managed-agents-2026-04-01 header.
Do I need a separate API call to start a workflow run?
No. You send a user message describing the work. The agent decides whether to start a run. Guide that choice in the system prompt.
How much does a Managed Agents session cost in USD?
Tokens at each model’s list rates, plus $0.08 per hour while the session status is running. A run has no flat fee of its own.
Can permission policies block a run from starting?
No. Policies apply to tools the run’s agents call. To stop runs entirely, disable workflows on the agent.
What are the main scale limits?
About 64 concurrent threads per run (not API-guaranteed), 1,000 agents started over a run’s life, a 24-hour default lifetime, and 10 open runs per session by default.
1 comment