How Anthropic’s Dynamic Workflows Let Claude Coordinate Its Own Agent Teams
Anthropic announced dynamic workflows at AI Engineer New York 2026, a feature that lets Claude autonomously plan, launch, and coordinate its own teams of subagents for complex tasks. The feature is available in Claude Console for API users, costs significantly more than single-agent approaches, and comes with honest constraints around budget visibility, agent-to-agent messaging, and self-improvement.
Claude Can Now Build and Manage Its Own Agent Teams
At AI Engineer New York 2026, Anthropic announced a feature that changes how complex AI tasks are structured: dynamic workflows, also referred to informally as “flows.” Latent Space’s coverage of the event — available at latent.space/p/claude-dynamic-workflows — documents how Anthropic’s CJ Avilla and Harrison Stall, both Members of Technical Staff, presented the feature and walked the audience through what it means for developers and ambitious AI-powered projects.
If you have been following the multiagent AI story — where multiple AI instances work together to handle tasks too large for a single conversation — this announcement is a meaningful step forward. Until now, setting up that kind of coordination required a developer to explicitly configure which agents would exist, what each one would do, and how they would hand off information. Dynamic workflows change that equation: Claude itself decides how to divide and run the work.
What Dynamic Workflows Actually Are
According to Latent Space’s discussion of the session, flows are defined by Anthropic as “a new type of multiagent orchestration, built for your most ambitious workloads.” The mechanism works in three stages. First, a lead Claude agent creates a plan for how it wants to accomplish the task at hand. Then, Claude writes an orchestration program — essentially a coordination script — that launches subagents, assigns them portions of the work, collects their results, and manages how information moves between them. Finally, the whole process produces an output without the developer needing to manually wire up every connection.
This is a departure from how Claude Managed Agents previously worked. Latent Space’s account of the presentation notes that previously, developers had to configure a coordinator and a roster of subagents themselves. With dynamic workflows, Claude determines that topology on its own.
Harrison Stall is quoted in the Latent Space coverage explaining it this way: flows are “the next version of multi-agent orchestration, where you can let Claude fully orchestrate the entirety of its own multi-agent topology.”
The analogy Avilla and Stall drew at the conference is instructive. They opened by invoking the history of Unix and Linux — Ken Thompson building Unix alone, then Linux evolving from Linus Torvalds to a layered system of trusted maintainers. The lesson they drew: as agents multiply and contribute more work in parallel, someone (or something) needs to act as a trusted coordinator. Dynamic workflows are Anthropic’s answer to that coordination problem at scale.
What This Enables — In Plain Terms
Consider a research and compliance team at a mid-sized mutual fund in Pune. The team tracks forty or fifty listed companies across sectors, reading quarterly results, regulatory filings, and analyst commentary. Today, a human analyst might spend a full day just on one sector. With dynamic workflows, a lead Claude agent could receive the instruction — “scan this watchlist and produce a risk summary” — and then autonomously spin up separate subagents for each company or sector, each working in parallel with its own isolated context. A later review phase then consolidates findings before producing the final report.
This is almost exactly the stock watchlist scanner that Avilla and Stall demonstrated at AI Engineer New York. Latent Space’s coverage describes how “a scan phase distributed individual tickers to agents with isolated contexts, and a later review phase examined the findings before producing a report.” For the Pune compliance team, the benefit is not just speed — it is the fact that no single agent is trying to hold fifty companies’ worth of data in one context window, which degrades quality.
You do not need to be a developer to understand why this matters. If you are a business analyst, a legal researcher, or a content strategist managing large, multi-part projects, the shift is this: instead of breaking your task into pieces yourself and running Claude repeatedly, you describe the goal once and the system figures out the division of labour.
Where Flows Live and How You Access Them
Dynamic workflows are available inside Claude Console, which Latent Space describes as the interface developers use when accessing Claude through Anthropic’s API — the home of Claude Managed Agents more broadly. Developers can also introduce specialist subagents of their own design into a flow if they choose; Avilla noted in the session that “it’s also possible in the agent config to pass in pre-configured sub-agents.”
For visibility into what is happening inside a flow, Claude Console includes a graphing functionality. Latent Space’s account of the Q&A records Avilla explaining that developers can inspect exactly what payload the lead agent passed to each subagent: “You can look at that payload here to deeply understand what is actually happening for every transition in this entire graph.”
Limitations and Tradeoffs You Should Know
Dynamic workflows are powerful, but the Latent Space coverage surfaces several honest limitations that matter before you get excited about replacing entire analyst teams.
The Bigger Vision Anthropic Is Working Toward
The most revealing moment in the Latent Space coverage comes from the Q&A, when an audience member asked whether dynamic workflows could power large scientific research — referencing OpenAI’s recently announced work on the Navier–Stokes Millennium Prize problem. Stall declined to discuss specific computations but was direct about the ambition: “Flows are basically a way to throw test-time compute at hard problems and get more capability faster out of the models. 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.”
That framing — scientific discovery accessible to general users, not just research labs — is worth holding in mind. It signals that dynamic workflows are not just a developer convenience feature. They are Anthropic’s infrastructure bet on what serious, long-horizon AI work looks like.
What to Watch For Next
If you are a non-technical professional in India curious about whether this affects you directly today, the honest answer is: not yet, unless your organisation is already accessing Claude through the API and working with a developer on Claude Managed Agents. Dynamic workflows live in Claude Console, which is an API-tier product.
What is worth watching: whether Anthropic brings flow-based orchestration into more accessible interfaces over time, and how the dreaming feature matures from research preview into a production-ready tool. For now, the feature represents a meaningful shift in how AI systems self-organise — one that is worth understanding even if you are a few steps removed from the technical implementation.
