On Friday, October 9, 2026, Anthropic added a new capability to Claude Managed Agents, its service for running artificial intelligence agents in the cloud: dynamic workflows, available in public beta. The feature lets a lead agent write a plan in the form of a program, run it in the background by distributing tasks among many agents, and then gather the results. According to the official documentation, a single run can start up to 1,000 agents.
Claude Managed Agents is Anthropic’s managed infrastructure for those who do not want to build an agent loop, its runtime and its tools from scratch. It offers isolated environments where Claude can read files, run commands, browse the web and execute code, with persistent sessions and scheduled runs. Until now, an agent could delegate to subagents step by step; what is new is that the entire coordination is written as a program.
The difference matters. In classic delegation, the lead agent decides at each moment what to hand off, waits for the report and moves on. With dynamic workflows, a program written by the agent itself defines the phases, launches the agents, collects what each one returns, passes it to others and repeats steps when needed. Meanwhile, the lead agent can keep talking with the user or end its turn and check on progress later.
Anthropic suggests uses such as audits, migrations, deep research and cross-checking sources. One example in the documentation is reviewing hundreds of documents: one agent per contract analyzes its contents and other agents review all the contracts again, not only the ones flagged in the first pass, so that mistakes do not slip through. If a review fails, the work is redone and checked again.
To make its case, the company published an internal test. Its team hid 70 bugs in a 116,000-line codebase and compared two approaches over three runs each. A single agent found 14, 15 and 27 bugs; the dynamic workflow found 66 every time, a detection rate of 94.3%. The outlet The Decoder noted that it is unclear whether those gains will hold for other kinds of tasks.
The technical limits qualify the eye-catching figure of a thousand agents. According to Anthropic’s documentation, 1,000 is the total number of agents a workflow can start over its whole lifetime, but only 64 threads work at once in a single run, and the company warns that this number is not guaranteed and may change. Each run lasts up to 24 hours by default, and a session allows at most ten open runs at the same time. If the workflow asks for more agents than allowed, the run ends with an error.
Cost is the other big question. The feature has no fee of its own, but each agent consumes tokens at the price of the model it uses, and Managed Agents also charges $0.08 per session hour. Anthropic warns that these workflows can use “a lot of tokens” and recommends starting with small jobs and setting a session budget, which is shared by all the agents in the workflow. In an earlier piece on multi-agent systems, the company had already noted that they typically consume 3 to 10 times more tokens than a single agent.
The debate over whether it is worth it remains open. According to The Decoder, a senior OpenAI engineer recently called agent swarms a massive waste of tokens. Anthropic’s test suggests the opposite for bug hunting, but it does not report total spending or false positives, so the efficiency comparison is incomplete. The general advice is to measure with your own workloads before adopting it at scale.
There are also design details worth knowing. The agents in a workflow share the files of the runtime environment, which spares the lead agent from loading huge volumes of results, but it can cause conflicts if several agents modify the same file. In addition, a run marked as completed does not guarantee that every assignment succeeded: the program can finish even if some agents fail, so the workflow itself must report what was left unreviewed.
To turn it on, developers must configure the multi-agent type that corresponds to the October 2026 version, with workflows enabled; with that type, both workflows and delegation to subagents are on by default. After that, it is enough to describe the work in a message or in the system instructions, and the lead agent decides whether to launch a workflow. Agents can be defined on the fly or picked from a list prepared in advance, each with its own tools and permissions.
The feature is not entirely new in Anthropic’s ecosystem: dynamic workflows were announced for Claude Code on May 28. What changes now is that they arrive in the managed cloud environment, with different limits and conditions, designed for long, asynchronous jobs that do not depend on a local computer.
The move adds to a clear industry trend: shifting from an assistant that answers questions to systems that split a large job among many specialized instances. Google introduced its “Gemini agent” this same week, capable of coordinating subagents, and OpenAI is pushing its own agent tools. Competition is no longer fought only on model quality, but on the infrastructure that lets companies organize, monitor and pay for the work of hundreds of agents at once.
It remains to be seen how the feature will behave outside controlled tests. The key questions are familiar: how much each completed task really costs, how many errors slip through, what happens when agents disagree and who reviews the final result. If dynamic workflows deliver what the 70-bug test promises, they could become a standard tool for audits and large-scale reviews; if not, they will serve as a reminder that more agents do not always mean better results.
