At the AI Engineer New York 2026 conference this morning, Anthropic engineers presented a new feature for Claude Managed Agents called dynamic workflows — referred to repeatedly in the sessions as “flows.” The talk was delivered by CJ Avilla and Harrison Stall, both Members of Technical Staff at Anthropic.
What are flows?
Anthropic describes flows as a new form of multi-agent orchestration designed for ambitious workloads. Rather than requiring developers to wire up a fixed coordinator and roster of subagents, flows let Claude determine how to organize and split work across agents: a lead agent creates a plan, Claude generates an orchestration program that launches subagents, aggregates results, and manages information flow.
The presenters compared this need for structured organization to historical operating-system projects — from Ken Thompson’s original Unix through Linux’s evolution from a single maintainer to layers of trusted maintainers — to emphasize that multi-agent systems similarly require organization to scale to larger projects.
How it works in practice
Flows are accessible through Claude Console, the interface used by developers accessing Claude via API, and are part of the Claude Managed Agents feature set. In a demo, the team showed a stock watchlist scanner as a simple example: a scan phase distributes individual tickers to agents running in isolated contexts, and a later review phase inspects findings and produces a report. The demo illustrated the core pattern, though Anthropic positions flows for much larger workloads.
Users can still supply pre-configured specialist subagents in the agent configuration if desired.
Budgets, parallelism and visibility
Audience questions highlighted practical considerations. When asked how Claude decides how many subagents a task needs — and whether a simple ticker scan could spawn hundreds of subagents — Stall said the answer depends on evals and default limits; ambitious parallelization generally requires explicit encouragement.
Controlling token/billing budgets is another core concern. Stall noted the model does not inherently know its budget while running, but developers can set a session budget. He added that telling models their budgets could make them "a little bit anxious," a remark that drew laughter.
The presenters emphasized that flows typically cost “much more” than a single-agent approach but can be much more effective at finding details.
Making flows more efficient: dreaming and memory
For repeated flows, the team suggested using Anthropic’s experimental “dreaming” feature. According to Anthropic’s documentation, a dream reads an existing memory store alongside past session transcripts and produces a reorganized memory store: merging duplicates, replacing stale or contradicted entries, and surfacing new insights.
Avilla described dreaming in the flow context as a way to crystallize and improve the memory store related to sessions, which could make repeated runs cheaper and more efficient over time.
Limits and controls
Other audience questions covered operational visibility and inter-agent communication. Avilla said developers can inspect subagent actions through Claude Console’s graphing tools, examining the payloads the lead agent passed to subagents to understand each transition in the graph. Stall added that subagents share a filesystem and can use it to exchange information, but developers cannot directly instruct one workflow subagent to message another via the API.
Intended scale and goals
When asked whether flows could power large scientific research projects — referencing OpenAI’s recent claim about the Navier–Stokes Millennium Prize problem — Stall declined to discuss specifics on how Anthropic computes large scientific problems. He did say Anthropic’s goal is to enable general users to tackle hard problems as well.
"Flows are basically a way to throw test-time compute at hard problems and get more capability faster out of the models," Stall said. "We think, especially as an API service, that unlocking the power of this is the first step for us making these types of discoveries, or scientific progress, broadly doable by the general public."
Summary
Anthropic’s dynamic workflows let Claude automatically design and orchestrate multi-agent topologies from a lead-agent plan, with management and visibility through Claude Console. The feature targets large, multi-phase workloads and research-style problems, but it increases compute costs and requires explicit settings to manage parallelism and budgets. Additional tools such as dreaming and session budgets are suggested paths to improve efficiency and maintain control.
Sponsor note: Union.ai appeared in conference materials promoting durability and runtime management for cloud-based workflows.
