How Anthropic's Dynamic Workflows Coordinate Claude Agents

At AI Engineer New York 2026, Anthropic presented dynamic workflows, announced on Friday and called flows by the presenters. latent.space's report quotes the definition: "a new type of multiagent orchestration, built for your most ambitious workloads." CJ Avilla and Harrison Stall, Members of Technical Staff at Anthropic, led the Claude Managed Agents workshop as the conference opened with morning workshops at AI Engineer NYC. Richard MacManus dated the account Oct 12, 2026.

This piece is for developers who call Claude through the API in Claude Console. It covers the move off a fixed coordinator and subagent roster, the orchestration program a lead agent writes, and the stock watchlist scanner demo. It also tracks agent count, session budgets, dreaming, graph inspection, filesystem sharing, cost, and scientific workloads.

Anthropic's Dynamic Workflows Coordinate: what actually changed

Avilla and Stall invoked Ken Thompson's original Unix operating system, then Linux moving from one person, Linus Torvalds, to layers of trusted maintainers. The lesson: agents need that organization for the ambitious projects Anthropic envisions for dynamic workflows. Avilla said people managing many agents "might start to feel this feeling of being the bottleneck and being the breaking point" amid "tons and tons of PRs," while "sitting there in our GitHub trying to approve PRs and make sure that everything works as expected."

Flows began in Claude Code as the "a harness for every task" concept and can run longer than a single agent or a roster. Claude chooses the organization of agents. Stall called them "essentially the next version of multi-agent orchestration," where "you can let Claude fully orchestrate the entirety of its own multi-agent topology." Claude now determines how to divide and orchestrate the work, instead of a required coordinator and subagent roster. The agent config can "pass in pre-configured sub-agents," Avilla said. This is available in Claude Console.

Anthropic's Dynamic Workflows Coordinate: how it works

The lead agent creates a plan for how it wants to accomplish tasks. Claude then writes an orchestration program, which launches agents, collects results, and manages information flow. Stall said that across phases it accomplishes "discrete, verifiable units of work" without the developer managing "any of the interconnections between these things." The stock watchlist scanner was far removed from the ambitious projects Anthropic is encouraging. A scan phase distributed individual tickers to agents with isolated contexts, and a review phase examined the findings before producing a report.

Managed Agents is located in Claude Console, alongside the watchlist scanner, a brain/hands conceptual overview of Managed Agents, and an image via Anthropic. The agents share a filesystem and can use it to exchange information, but developers cannot directly instruct one workflow subagent to message another through the API. Avilla said this is visible through graphing inside Claude Console. "You can tell, what did the lead agent actually pass to the sub-agent? You can look at that payload here to deeply understand what is actually happening for every transition in this entire graph."

Anthropic's Dynamic Workflows Coordinate: why it matters now

Anthropic used lessons from Unix and Linux to explain how agents can tackle bigger projects without overwhelming developers. The strain is a person in GitHub approving agent pull requests and checking that the work behaves. Flows suit workloads past a single agent or a roster, with Claude choosing the topology. In the watchlist-scanner comparison, Avilla said flows cost "much more" than a single-agent approach and were "much more effective at finding all of the details." Beside that comparison: "Setting a budget is key!"

An audience member asked whether dynamic workflows can power large scientific research projects, citing OpenAI's recently announced proof on the Navier–Stokes Millennium Prize problem. Stall would not be drawn on specifics about how Anthropic computes those problems, but said the goal is for general users to tackle them too. "Flows are basically a way to throw test-time compute at hard problems and get more capability faster out of the models," he said. Especially as an API service, he called unlocking the power of this the first step toward making these types of discoveries, or scientific progress, broadly doable by the general public.

Anthropic's Dynamic Workflows Coordinate: who is affected

The developers affected use Claude Managed Agents in Claude Console, the API interface and the whole point of that product. They used to configure a coordinator and a roster of subagents. Claude now determines how the work is divided, and the agent config can still pass in pre-configured sub-agents. A watchlist run gives each ticker an isolated context in the scan phase, then a review phase produces a report. The heavier case is the GitHub reviewer Avilla called the breaking point.

One person asked how Claude decides how many agents a task needs, and whether a simple ticker scan could trigger hundreds of sub-agents. Stall replied, "The answer really is evals and lower defaults," and said ambitious tasks generally require explicitly encouraging greater parallelization. Asked whether the lead agent knows how much money it has available, he said the model does not know its budget while running. Developers can set a session budget. Informing models of their budgets could make the agents "a little bit anxious," which drew chuckles.

Anthropic's Dynamic Workflows Coordinate: what to watch

Developers also asked how to make repeated flows cheaper and more efficient over time. Outside of asking Claude to examine the costs incurred, Avilla pointed to dreaming, a relatively new research preview in the documentation. A dream "reads an existing memory store alongside past session transcripts, then produces a new, reorganized memory store," merging duplicates, replacing "stale or contradicted entries," and surfacing "new insights." In a flow, he said, dreaming can "crystallize and basically improve the memory store that is related to those sessions."

Could the watchlist flow recognize an inadequate scanning model and change the next scheduled run? Stall said this was not built in. He and Avilla suggested hill-climbing and dreaming could be used to improve the loop. Sponsor Union.ai, at AI Engineer New York booth #48, described a sovereign AI factory where engineering teams own data, compute, infrastructure, and a durable runtime on their cloud. Workflows adapt, right-size infra resources, and recover from failures without re-running completed work, free for 30 days through AWS Marketplace.