Claude writes the orchestration program
Latent Space reported Anthropic's talk at AI Engineer New York on October 12. Members of technical staff Harrison Stall and CJ Avilla described dynamic workflows, which they also called flows. Previously, Claude Managed Agents required developers to configure a coordinator and a roster of subagents. With dynamic workflows, Claude itself determines how to divide and orchestrate the work.
The account of the mechanism is: the lead agent creates a plan, then Claude writes an orchestration program, which launches agents, collects results, and manages the flow of information. Developers can still pass in pre-configured sub-agents. Flows sit in Claude Console, the interface for developers using Claude through the API.
[1]
The demo was a stock watchlist
The example on stage was a stock-watchlist scanner. A scan phase distributed individual tickers to agents with isolated contexts, and a later review phase examined the findings before producing a report. Latent Space writes that this was far from the large projects Anthropic is encouraging. It illustrated the basic pattern.
In that comparison, Avilla said flows cost much more than a single-agent approach and were much more effective at finding all of the details. The article gives no multiplier for the cost and no score for the result.
[1]Budget, communication, and self-improvement all have limits
An audience member asked how Claude decides how many agents a task needs, and whether a simple scan could launch hundreds. Stall's answer was evals and lower defaults. Ambitious tasks generally require explicitly encouraging more parallelization.
The model does not know its budget while running. Developers can set a session budget. Stall said that telling models their budgets could make the agents a little anxious. Subagents share a filesystem and can use it to exchange information, but developers cannot directly instruct one workflow subagent to message another through the API. Graphs in the Console show what the lead agent passed to a subagent.
Can the watchlist notice that its scanning model is inadequate and change the next run? Stall said that is not built in. The two suggested hill-climbing and dreaming as ways to improve the loop. In the documentation, dreaming is a research preview: it reads an existing memory store and past session transcripts, then produces a reorganized memory store. That is a suggested use, not something the demo already does.
[1]