Anthropic’s dynamic workflows — “flows” — are a new type of multi-agent orchestration in which Claude, not the developer, decides how to organise the work: a lead agent writes a plan, then produces an orchestration program that launches subagents, collects their results and manages the information flow between them. Members of Technical Staff CJ Avilla and Harrison Stall explained the feature at AI Engineer New York on October 12, 2026, three days after Anthropic announced it, and it is available in Claude Console.
Key facts
| Fact | Detail |
|---|---|
| Question answered | Anthropic’s Dynamic Workflows Let Claude Choose Its Own Agent Roster |
| Short answer | Anthropic used its AI Engineer New York session on October 12, 2026 to explain dynamic workflows, a multi-agent orchestration feature announced the previous Friday. |
| Reading time | ~4 min |
| Last updated | 2026-10-13 |
What changed
- The roster is no longer yours to wire up. Claude Managed Agents used to need a coordinator plus a hand-configured set of subagents. Flows invert that: the model picks the topology.
- A plan, then a program. The lead agent plans the task, and the orchestration program it writes launches agents, gathers results and routes information between phases.
- Structured phases still exist. In the demo, a stock watchlist scanner split into a scan phase that handed individual tickers to agents with isolated contexts, and a review phase that examined the findings before writing a report.
- You can still pin specialists. Pre-configured subagents can be passed in through the agent config.
- Inspecting the work is possible. Subagent activity can be reviewed through graphing functionality in the product.
Why Anthropic is doing this
The framing at the session was Unix and Linux: Ken Thompson’s original operating system, then Linus Torvalds and layers of trusted maintainers. The parallel is not decorative. Avilla described the failure mode developers already feel — running many agents means many pull requests, and the human sitting in GitHub approving them becomes “the bottleneck and the breaking point” of the whole arrangement.
Flows are an attempt to push that coordination work down into the system. If Claude can plan the decomposition and run discrete, verifiable units of work across phases without anyone managing the interconnections, then the number of agents stops being limited by how many dashboards one person can watch.
The questions the audience actually asked
Three answers from the Q&A are worth keeping, because they are the parts vendors usually leave out.
- Agent count is a defaults problem, not a capability problem. Asked whether a simple scan could trigger hundreds of subagents, Stall answered “evals and lower defaults”: ambitious fan-out has to be asked for, rather than being the baseline behaviour.
- The model does not know its budget. Developers can set a session budget, but the running agent is not told what it has. Anthropic’s lighter note — that informing models of their budgets could make them “a little bit anxious” — still leaves you with the operational question of who watches spend if the orchestrator does not.
- Self-improvement is not built in. When asked whether a flow could notice its own scanning model was inadequate and change the next scheduled run, Stall said no. Anthropic points at two adjacent mechanisms instead: hill-climbing and “dreaming”, a research-preview feature that reads an existing memory store alongside past session transcripts, merges duplicates, replaces stale or contradicted entries and surfaces new insights.
What this means if your agents run on your own machine
There are two orchestration problems, and only one of them is solved by a smarter coordinator. Deciding how to split a project across agents is the half Anthropic is working on. Executing each unit of work against real files, real commands and a real desktop is the other half, and it is still local — which is why an approval-first agent that runs natively on Windows, proposes a plan and waits for your sign-off before it touches anything is worth having regardless of how good the planner gets. MOOGH is that half, and it is a free download.
What to watch next
- Whether orchestration quality holds up outside demos on tasks with genuinely ambiguous scopes.
- How budgets are surfaced and capped once flows run on schedules rather than in sessions.
- Whether the graphing view is enough to audit what a fan-out of subagents actually did.
FAQ
- Do I need Managed Agents to use flows? Flows are part of the Managed Agents surface and are used through Claude Console.
- Is this only for large projects? The session argued it is built for ambitious workloads, and the demo was deliberately modest — a watchlist scan — to show the pattern rather than the scale.
- Does the lead agent write code I have to maintain? It writes an orchestration program for that run; the plan and the phases are the artefacts you inspect, not a codebase you own.
Related reading: