Anthropic put dynamic workflows into Claude Managed Agents in public beta on 9 October 2026. An agent can now write a workflow, a program that runs many agents in phases and combines their results, and Anthropic’s documentation caps the agents one workflow run starts at 1,000.
The @ClaudeDevs launch post on X had about 870,000 views when read on 11 October. In a reply, the account gave Anthropic’s own test: across three runs on a codebase with 70 planted bugs, a single agent found 14, 15 and 27, while a workflow found 66 each time.
The release notes describe it in one line: “For work with many pieces, such as reviewing hundreds of documents, the agent can write a workflow.” The server runs that program in the background as a workflow run, and developers follow it through workflow_run events on the session’s event stream.
It is in beta behind the managed-agents-2026-04-01 beta header. To turn it on, the agent’s multiagent field is set to the multiagent_20261001 type with workflows enabled; the docs say workflows and subagents are both on by default with that type.
The launch post called it “a new type of multiagent orchestration, built for your most ambitious workloads.” It summed up the mechanism as: “A lead agent writes a plan that runs across many agents in phases, combining the results at the end.”
Bugs found out of 70 planted in a 116k-line codebase, three runs each, as posted by @ClaudeDevs on 9 October 2026. The test is Anthropic’s own; the post does not give the model, token use or cost per run.
Business Pill · MULTI-AGENT SYSTEMS
A one-minute explainer of multi-agent systems: how several agents split one job and combine their results. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, Anthropic is moving orchestration from the developer’s code into the agent itself: the agent writes the program that fans work out to up to 1,000 agents, and the platform bills it as ordinary token use plus session runtime.
How a Run Works
Anthropic’s docs contrast it with subagents: “With dynamic workflows, Claude writes a program to orchestrate agents without Claude’s direct involvement.” With subagents, the agent delegates tasks itself and reads what each one reports.
No extra API call starts a run. The docs say the developer describes the work in a message, and the agent decides whether and when to start a run, guided by its system prompt.
The agents in a run can be defined by the workflow itself or picked from up to 20 agents the developer lists. The docs list audits, migrations, deep research and cross-checking as examples of the work it suits.
The Limits in the Docs
The workflow runs page gives four limits. A run can have 64 threads working at once, and the docs say the API does not guarantee that number. A workflow can start 1,000 agents over a run’s whole life; past that, the run ends with a thread_limit_error.
A run lasts 24 hours by default, or the lifetime the agent sets. A session can hold 10 open runs by default, and idle runs count toward that.
The docs add that “The server has other limits on workflows, and rules for them, that aren’t listed here.”
Limits as listed on Anthropic’s “Workflow runs” docs page, read 11 October 2026; the feature is in public beta.
What It Costs
“A run has no price of its own,” the docs say. The tokens its agents use are billed like the session’s other tokens, at each model’s rates, and Claude Managed Agents also bills session runtime at $0.08 per session-hour while a session is running.
The docs repeatedly warn about spend. They say every agent in a run uses tokens, recommend a session budget to cap the session’s spend, runs included, and note that at the budget every open run pauses.
Anthropic’s Bug-Hunt Test
The figures come from the @ClaudeDevs reply: “We planted 70 bugs in a 116k-line codebase. Across 3 runs, a single agent found 14, 15 and 27 bugs. A workflow consistently found 66 in each of its 3 runs.”
The same approach reached Claude Code first. A claude.com post dated 28 May 2026 introduced dynamic workflows there, saying Claude “dynamically writes orchestration scripts that run tens to hundreds of parallel subagents in a single session.” That post has since been updated to say the feature is generally available.
The May post gives one example: it says Bun’s Jarred Sumner used dynamic workflows to port Bun from Zig to Rust with 99.8% of the existing test suite passing, roughly 750,000 lines of Rust, in eleven days from first commit to merge.
The Structural Read
In the docs, the split between the two modes is who drives. With subagents, the agent delegates and reads each report; with dynamic workflows, a program the agent wrote does the coordinating, and context and results pass between agents programmatically.
The costmodel follows from that design. A run has no separate price, so its spend scales with how many agents the workflow starts and how many tokens each uses, which is why the docs pair the feature with session budgets.
“A lead agent writes a plan that runs across many agents in phases, combining the results at the end.”
Three Implications
TEAMS BUILDING ON MANAGED AGENTS The docs say the agent decides when to start a run, guided by its system prompt, so the prompt becomes where a team sets when fan-out is worth the tokens.
FINANCE AND PLATFORM OWNERS Runs are billed at each model’s token rates plus $0.08 per session-hour of runtime; the docs recommend a session budget, at which every open run pauses.
ANYONE WEIGHING THE BUG-HUNT FIGURES The 66-of-70 result is Anthropic’s own test; the reply does not give the model, the token use or the cost per run.
The essay argues that “The surface, not the model, governs autonomy”, and puts it more bluntly: “The harness is the building.”
As we read it, dynamic workflows are a harness feature in exactly that sense: the same models, with the orchestration layer now written by the agent and run by the platform.
What Is Not Established
The bug-hunt figures are Anthropic’s own test, given in a short reply on X. The reply does not say which model ran, how many tokens or dollars each run used, or how many of the reported bugs were real finds versus false positives.
The Managed Agents feature is a beta. The docs say the 64-thread figure can change and that other workflow limits are not listed.
Business Engineer Framework
The Agentic Harness War
A Business Engineer essay on why the agentic surface, not the model, is becoming where value pools in the AI stack.
Claude Managed Agents can now let an agent write its own orchestration program and run up to 1,000 agents over a single workflow run, in public beta since 9 October 2026. Anthropic prices the runs as ordinary token use plus $0.08 per session-hour, and points to its own test in which a workflow found 66 of 70 planted bugs in each of three runs against 14 to 27 for a single agent.
A note on sourcing. We read Anthropic’s 9 October release notes, its documentation pages on multiagent orchestration, workflow runs and Managed Agents pricing, its 28 May Claude Code blog post and the @ClaudeDevs posts on X in full on 11 October 2026. Every figure here is Anthropic’s own; we have not run the bug-hunt test ourselves. Nothing here is a forecast, and nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
