At AI Engineer New York on Monday, CJ Avilla and Harrison Stall from Anthropic ran a workshop on Claude Managed Agents. The centerpiece was the feature Anthropic announced on Friday: dynamic workflows, which both presenters kept calling “flows.” I covered the announcement itself on Friday; what the workshop adds is how the thing actually works.
Claude writes the orchestration program
Until now, Managed Agents meant configuring a coordinator and a roster of subagents yourself. With flows, Claude decides how the work gets split. The lead agent builds a plan and turns it into an orchestration program: it launches agents, collects results, and manages what information goes where.
“Flows are essentially the next version of multi-agent orchestration,” Stall said, “where you can let Claude fully orchestrate the entirety of its own multi-agent topology.” Across several phases, it handles discrete, verifiable units of work without you managing the connections between them. Your own specialists still fit: pre-configured subagents can be passed in through the agent config.
Their way in was, of all things, Unix and Linux, from Ken Thompson’s original to Torvalds’ layers of maintainers. The point behind it is less grand than it sounds. Avilla described the feeling of becoming the bottleneck yourself, sitting in GitHub approving pull requests from a dozen agents.
The Q&A was the interesting part
How does Claude know how many agents a task needs, and could a simple ticker scan spiral into hundreds of subagents? Stall’s answer: evals and lower defaults. Ambitious tasks need you to explicitly ask for more parallelization.
Does the lead agent know its budget? No. The model doesn’t know how much money it has while running, Stall said; what you can set is a session budget. Telling models their budgets could make the agents “a little bit anxious” — good for a laugh in the room, and still a useful hint.
How do repeated flows get cheaper over time? Avilla pointed at “dreaming,” a research preview feature per the documentation: it reads an existing memory store alongside past session transcripts and produces a reorganized store, duplicates merged, stale or contradicted entries replaced, new insights surfaced.
Two more answers matter in practice. Everything is inspectable: a graph view in Claude Console shows, for every transition, what the lead agent handed to which subagent. And subagents can’t talk to each other directly. They share a filesystem and exchange information through it, but the API gives you no way to tell one workflow subagent to message another.
Flows are compute converted into capability
Avilla says outright that flows cost much more than a single-agent approach, using the same stock-watchlist example. In exchange they are much better at finding all the details. The flows can’t improve themselves, either: a scheduled run won’t notice that its scanning model is inadequate and won’t change course.
Which leaves the question of what this is for. Stall is direct about it: flows are basically a way to throw test-time compute at hard problems and get more capability out of the models, with the goal of making scientific progress broadly doable by the general public. That’s a big claim for a feature whose demo was a ticker scanner. It does explain why the budget is a dedicated knob rather than a footnote.