Anthropic’s dynamic workflows are a new orchestration layer for Claude Managed Agents, and the company’s own framing is the useful part: “a new type of multiagent orchestration, built for your most ambitious workloads.” Announced the Friday before the talk and presented at AI Engineer New York by CJ Avilla and Harrison Stall, both members of technical staff, the feature is aimed at a problem anyone running several agents already recognises - the human in the middle. Source: https://www.latent.space/p/claude-dynamic-workflows
Key facts
| Fact | Detail |
|---|---|
| Question answered | Anthropic’s Dynamic Workflows: Running Claude Agents Like an Operating System |
| Short answer | Anthropic used AI Engineer New York to present dynamic workflows, a multi-agent orchestration layer for Claude Managed Agents that the company calls ’a new type of multiagent orchestration, built for your most… |
| Reading time | ~5 min |
| Last updated | 2026-10-17 |
The Unix analogy is not decoration
Avilla and Stall opened with operating-system history: Ken Thompson building the original Unix, then Linux growing from one maintainer into layers of trusted maintainers. The lesson they drew is that big projects do not scale on heroic individuals; they scale on structure - ownership, interfaces, and a way to hand work between layers without re-litigating everything.
That is the argument for flows. Agents can already write code; what they cannot do is keep fifty parallel workstreams coherent without someone arbitrating. Anthropic’s answer is to put a lead agent in the position of a scheduler and let subagents do bounded work underneath it.
“Being the bottleneck and being the breaking point”
The most quotable moment was Avilla describing what happens when orchestration is missing:
“I think a lot of us who are orchestrating and managing lots of agents now might start to feel this feeling of being the bottleneck and being the breaking point as we have lots of agents contributing tons and tons of PRs. And we’re sitting there in our GitHub trying to approve PRs and make sure that everything works as expected.”
That is a precise description of the failure mode. The agents get faster; the review queue does not. Adding more agents to a workflow whose bottleneck is human approval just makes the queue longer.
How a flow is actually wired
From the talk and the Q&A, the shape is:
- A lead agent decomposes the task and decides how much parallelism to use. Asked whether a trivial trigger could fan out into hundreds of subagents, Stall answered: “The answer really is evals and lower defaults,” adding that ambitious tasks generally need explicit encouragement to parallelise.
- Subagents with bounded jobs. In the demoed stock-watchlist scanner, the work is split across scanning, checking and reporting rather than handed to one agent in one pass.
- A shared filesystem. Subagents exchange data through files. This is also the limit: Stall said there is no API for making one subagent message another directly.
- A graph you can inspect. In Claude Console you can see what the lead agent passed to each subagent. Avilla: “You can look at that payload here to deeply understand what is actually happening for every transition in this entire graph.”
What flows cost, and the budget question
Anthropic is explicit that this is the expensive path. In the watchlist-scanner comparison, Avilla said flows cost “much more” than a single-agent approach but were also “much more effective at finding all of the details.”
The budget mechanics are worth knowing before you build one:
- The model does not know how much money it has available while it runs.
- Developers set a session budget instead.
- Stall noted that telling models their budget could make agents “a little bit anxious” - a joke with a real point: budget is a control surface for the operator, not a planning input for the model.
“Dreaming” as the memory maintenance loop
Asked how to make repeated flows cheaper over time, Avilla pointed at dreaming, a research preview. Per Anthropic’s 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 flow terms, that is housekeeping: the memory the subagents rely on gets compacted between runs instead of growing until it is useless. Avilla described it as being able to “crystallize and basically improve the memory store that is related to those sessions.”
What flows do not do is improve themselves. Asked whether the watchlist flow could notice its scanning model was inadequate and change the next scheduled run, Stall said that was not built in; he and Avilla suggested hill-climbing and dreaming as the pieces you would combine to get there.
What this means if your agents run on Windows
Most readers of this site are not running Claude Managed Agents at scale. The pattern still transfers, because the bottleneck transfers:
- Steal the lead/worker split. Even with a desktop agent, one long prompt that does everything is the single-agent version of the same mistake. Give each step a bounded job and a defined output.
- Put the handoff on disk. Flows pass data through a filesystem because that is inspectable. On Windows, a working folder with numbered artefacts beats passing state through chat history - you can open the files, diff them, and rerun one step.
- Budget before you parallelise. Set a ceiling per run. The Anthropic position - the model should not be reasoning about its own budget - is a good default for local runs too.
- Keep the human at the decision, not the keystroke. The talk’s framing is that review should catch what checks cannot, not that review should disappear.
If what you actually need is an agent on the machine where the files live - installers, Office documents, local builds, Windows tooling - that is the other half of the picture: [download MOOGH] runs locally with per-action approval instead of shipping your work to a sandbox.
FAQ
Is dynamic workflows generally available? Anthropic announced the feature the Friday before the AI Engineer New York talk and presented it as part of Claude Managed Agents. Check Anthropic’s own documentation for current availability rather than relying on conference coverage.
Do I need Managed Agents to use the pattern? No. The lead-agent/subagent split, file-based handoffs and explicit budgets are design choices you can implement around any agent runtime, including a local one.
Why not just run one bigger model? Because a single pass misses details on large jobs - the presenters’ own comparison found flows more effective at finding everything, at a higher price. The trade is coverage against cost, and it only pays when the task is genuinely big.
Related reading: