Anthropic is letting Claude write the software that coordinates a team of other Claude agents, and the company’s own presenters say the approach costs much more than using a single agent. The feature is called dynamic workflows, or “flows,” and it was shown at the AI Engineer New York conference by CJ Avilla and Harrison Stall, both Members of Technical Staff at Anthropic, according to Latent Space.
Splitting a job across several agents creates a second job nobody asked for: deciding who does what, moving information between them, and checking what comes back. Before flows, a developer building on Claude Managed Agents, Anthropic’s hosted service for running agents, had to define the lead agent and list its helper agents by hand. With flows, a lead agent drafts a plan and then writes an orchestration program that launches the agents, collects their results, and manages how information moves. Stall said it lets Claude “fully orchestrate the entirety of its own multi-agent topology.” Developers can still pass in pre-configured subagents if they want specialists.
The motivation Avilla gave was human. People supervising many agents, he said, end up as the bottleneck, sitting in GitHub approving a flood of pull requests. Handing the coordination to the model is meant to remove that person from the middle.
The demonstration was a stock watchlist scanner. In the first stage, each ticker went to its own agent working in isolation. A second stage went over what those agents found before the report was written. Latent Space notes that the presenters themselves described it as far removed from the ambitious projects Anthropic has in mind. It was a walkthrough, not a test, and a single demonstration on stock tickers says almost nothing about how reliably flows work on real jobs.
The most useful admission came in the comparison with a single agent. Avilla said flows cost “much more,” and were also “much more effective at finding all of the details.” The source reports no figure for either side of that trade, so the size of the premium is unquantified.
The second gap concerns money. On whether the lead agent is told what it can spend, Stall’s answer was no: the running model has no view of its budget, although a developer can cap what a session costs. He added that telling models their budgets could make agents “a little bit anxious,” which drew laughs. Taken as described, the limit works as a ceiling the system runs into, not a number it can plan around. How many agents a task gets is steered instead by defaults: asked whether a simple ticker scan could spawn hundreds of subagents, Stall answered “evals and lower defaults,” and said ambitious tasks generally need explicit encouragement to parallelize.
Two more constraints came out of the questions. Subagents share a filesystem and can pass information through it, but the API gives a developer no way to order a subagent in a flow to send a message to a peer. And the Claude Console, the developer interface where Managed Agents lives, includes a graph view that shows what the lead agent passed to each subagent at every step.
Status is only partly clear. Latent Space calls flows a brand new feature announced the previous Friday and says it is available in the Claude Console. The report does not say whether it is a preview or generally available, and it gives no pricing. The one feature it labels a research preview is a separate memory tool called dreaming.
This is the vendor’s side of a trade we have covered before. Our earlier piece reported a study in which teams of AI coders cost up to 5.1 times more and rarely scored higher, and Anthropic’s own data showing quality flattening past ten agents while token costs kept climbing. Flows is not evidence for or against those findings, since Anthropic reported neither quality nor cost numbers here, but it shows the company building the thing that makes the trade easy to enter.
Anyone piloting flows should run the same task through a single agent first and set the session budget before the first run, because Anthropic has not published how large the premium is.
Reported by Richard MacManus for Latent Space, from the AI Engineer New York conference (the source page carries no publication date).
