Anthropic put Claude Dynamic Workflows into public beta on October 9, 2026. The feature sits inside Claude Managed Agents. A lead agent can write a background program that runs many helper agents in phases, then merges their results. For India and USA agencies building client AI systems, that is a concrete shift from one chatty agent to batch orchestration you can budget and observe.

This guide explains what shipped, what Anthropic’s own bug test showed, and how delivery teams should pilot it without burning tokens or client trust. Facts below come from Anthropic’s multi-agent orchestration docs and secondary reporting that quotes the same launch notes.

What Claude Dynamic Workflows actually do

Managed Agents already supported subagents. The lead agent delegates work, reads reports, and can send follow-ups. Claude Dynamic Workflows add a second mode. The lead writes a workflow: a program that schedules agents across named phases. Anthropic’s server runs that program in the background. The main session stays free to talk with the user and report progress.

You do not fire a separate “start workflow” API call. You describe the work. The agent decides whether a run is needed. Anthropic tells builders to put clear when-to-run rules in the system prompt. Small jobs stay on the lead. Large parallel jobs start a run.

Diagram comparing Claude Dynamic Workflows and subagents for agency orchestration — Claude Dynamic Workflows vs subagents. Debugged Pro original diagram based on Anthropic docs.

Subagents: lead can follow up; up to 25 child threads; one delegation level.
Dynamic workflows: lead writes phases; up to 1,000 agents over the run; up to 64 concurrent; no follow-ups after start.
Advisor (optional): primary thread can consult a stronger model mid-turn for planning or review.

Enable the multi-agent type multiagent_20261001 under the managed-agents-2026-04-01 beta header. With that type, workflows and subagents are on by default. You can disable either setting when you want a narrower agent.

Hard limits agencies must design around

Anthropic’s workflow runs documentation lists clear caps. Treat them as product constraints, not marketing fluff.

Limit	Value
Agents started over one run	1,000
Agents working at once	Up to 64 (not a hard guarantee)
Default run lifetime	24 hours (or shorter if the agent sets it)
Open runs per session	10 by default
Predefined agents listed for a workflow	Up to 20

The 1,000 cap counts agents started, not threads. The server can retry a failed agent on a new thread, so a run may show more than 1,000 threads. Going past the start cap ends the run with a thread limit error. Plan retries and partial coverage before you promise a client a full sweep.

Anthropic’s bug test: what it shows and what it does not

In the launch thread, Anthropic shared one internal test. The team planted 70 bugs in a 116,000-line codebase. Across three runs, a single agent found 14, 15, and 27 bugs. A workflow found 66 bugs in each of its three runs. The Decoder and other write-ups repeated those numbers from the same announcement.

What that helps with: large parallel search across many files benefits from fan-out. Variance dropped while coverage rose in that specific setup.

What it does not prove: it is Anthropic’s codebase and Anthropic’s planted bugs. No public cost figure came with the result. A senior OpenAI engineer has publicly called agent swarms a heavy token spend. Your agency still needs its own bake-off on a client-like repo before you sell “66 of 70” style outcomes.

For more on why extra agent code does not always mean more shipped software, see our related guide on the AI coding agents review bottleneck. Human review still absorbs a lot of the gain.

Cost model: no run fee, but tokens stack fast

A workflow run has no separate fee. Every agent bills tokens at its model rates. Managed Agents also charges session runtime at $0.08 per session-hour while the session is running. That meter is easy to miss when dozens of agents stay busy for hours.

Anthropic’s advice is blunt: start with a scoped task because dynamic workflows can use a lot of tokens. Session budgets are the brake. When list cost hits the budget, open runs pause. Each working agent finishes the request it already started, so spend can overshoot by one request per active agent. Raise or remove the budget only with an explicit approval path.

Set a session budget when you create the session. You cannot add one later.
Log controller tokens, worker tokens, and session-hours separately.
Compare single-agent cost and quality against a workflow on the same task.
Refuse open-ended “keep hunting” prompts without a time and dollar stop.

Why India and USA agencies should care now

Agencies in India and the USA sell delivery, not demos. Clients ask for audits, migrations, research packs, and cross-checks across hundreds of files or tickets. Those jobs match the workflow shape: many similar pieces, parallel workers, one merged report.

They also raise governance questions. Unsupervised agents that can hit the open web have already caused real-world messes. Our piece on unsupervised AI agents and false tips covers why containment matters. Claude Dynamic Workflows run inside Managed Agents with sandbox, vault, and MCP scoping, but you still must limit tools and hosts per role.

Debugged Pro already helps teams design agentic AI development and AI automation with budgets, evals, and human gates. The same discipline applies here. A bigger swarm is not a substitute for a clear RACI on merges, client data, and incident response.

Six-step agency pilot checklist for Claude Dynamic Workflows — Agency pilot checklist for Claude Dynamic Workflows. Debugged Pro original.

Pilot plan for Claude Dynamic Workflows

Use a two-week pilot. Keep the client logo off the first run if you can. Measure before you scale.

Week 1: scoped internal bake-off

Pick one batch job: contract pack review, static audit across 200 files, or migration checklist generation.
Create a lead agent with multiagent_20261001 and a tight system prompt for when to start a run.
Disable inline agents if you need strict tool control. List predefined agents only.
Set a low session budget. Force the team to watch pause events.
Run three single-agent baselines and three workflow runs on the same corpus.

Week 2: delivery packaging

Write a one-page cost/quality report for account managers.
Map controls to NIST AI RMF language clients already know: Govern, Map, Measure, Manage.
Define which outputs need human sign-off before they leave the agency.
Document MCP hosts, vault credentials, and allowed networks for limited environments.
Decide whether advisor models are allowed, and whether advice can stay redacted from clients.

If you need a broader consulting partner for architecture and rollout, start from our software development consulting page or browse more delivery notes on the Debugged Pro blog.

Prompt patterns that keep runs useful

Anthropic’s contract-reviewer example is a good template. Review one or two contracts yourself. When the stack grows, start a workflow that reads them in parallel and merges findings. You can add rules such as:

One agent per item, then a second pass that checks every item, not only the flagged ones.
One agent failure must not fail the whole run; list uncovered items clearly.
Hard time limit on the run, for example one hour.

Do not ask a workflow to “explore freely” across production systems. That is how agent demos turn into incident tickets. Keep tool lists short. Prefer read-only MCP where possible. Archive runs after reporting so leftover threads do not confuse the next session.

How Claude Dynamic Workflows fit with other agent work

Meta’s recent work on agentic meta-reasoning focuses on how a controller spends compute during long runs. Anthropic’s workflow feature focuses on how many agents can share a programmed plan. You may use both ideas: a controller decides whether a batch run is worth the budget, then a workflow executes the batch. See our agentic meta-reasoning agency guide for the compute-control angle.

Production trust is also falling for unreviewed agent changes, according to industry surveys covered on our site. Pair any multi-agent pipeline with mandatory human review for customer-facing or production edits. Speed is worthless if account teams spend the savings on apology calls.

Primary sources: Anthropic’s multi-agent orchestration docs for Managed Agents, plus The Decoder’s report on the 1,000-agent orchestration launch and CellCog’s summary of the October 9 public beta limits.

Key Takeaways

Claude Dynamic Workflows entered public beta on October 9, 2026 inside Claude Managed Agents.
A lead agent writes a phased program; the server can run up to 1,000 agents over a run, with up to 64 concurrent.
Anthropic’s internal bug test found 66 of 70 planted bugs per workflow run versus 14 to 27 for a single agent.
There is no separate run fee; tokens and $0.08 per session-hour still stack, so set session budgets first.
India and USA agencies should pilot on scoped batch jobs, measure variance and cost, and keep humans on release risk.

FAQ

What are Claude Dynamic Workflows?

They are a Claude Managed Agents feature. A lead agent writes a program that runs many agents in phases and combines results while the server executes the run in the background.

How do I turn on Claude Dynamic Workflows?

Use multiagent type multiagent_20261001 with the Managed Agents beta header. Workflows are enabled by default with that type. Guide the agent in the system prompt on when to start a run.

How is a workflow different from subagents?

Subagents let the lead follow up on each helper. A workflow runs as a program of phases without that back-and-forth. Once a run starts, the lead cannot message the run’s threads.

Can Claude Dynamic Workflows blow the budget?

Yes. Anthropic warns they can use many tokens. Set a session budget so runs pause at the cap. Start with a scoped task before client work.

Should India USA agencies use this in production this week?

Use a controlled pilot first. Prove quality and cost on an internal corpus. Keep human approval on merges, client data access, and any action that leaves the sandbox.

Nitin Kumar is Lead Project Coordinator at Debugged Pro. He works with India and USA agencies on AI delivery systems, agent guardrails, and practical rollout plans that survive client review.