You hand a project to one AI assistant and it’s done in minutes. You hand it something bigger, and suddenly you need five assistants, a coordinator, and someone keeping track of who passed what to whom. Anthropic just changed who handles that coordination job.
The company announced “dynamic workflows” (they call them “flows”) at the AI Engineer New York conference on October 12. The core idea: instead of a developer manually configuring which AI agents handle which pieces of a task, Claude figures that out on its own. It writes the orchestration program, spins up the agents it needs, and manages the information moving between them.
That’s a meaningful shift. Previously, building a multi-agent AI system meant specifying a coordinator and a roster of subagents upfront. You had to decide the structure before the work started. With flows, Claude assesses the task and designs its own team on the fly.
Anthropic demonstrated this with a stock watchlist scanner. Claude distributed individual tickers to separate agents running in parallel, then had another agent review all the findings before producing a final report. The system found more detail than a single-agent approach. It also cost more.
That cost tradeoff is honest and worth sitting with. More agents means more computation, more tokens burned, and higher bills. Anthropic engineers acknowledged during Q&A that the running model doesn’t know its budget while it’s working. Developers can set a session spending limit, but the agents themselves operate without that constraint in view. Think of it like giving a contractor a credit card: the card has a limit, but the contractor doesn’t see it while buying materials.
There’s also a feature called “dreaming,” which lets Claude review past session transcripts and reorganize its memory store, merging duplicates and surfacing new insights. A repeated workflow can get smarter over time without someone manually curating what the AI remembers.
For businesses running complex, multi-step analysis work (financial research, legal document review, software development, content pipelines) this kind of self-organizing AI infrastructure changes what’s practical to automate. Tasks that previously required carefully human-designed workflows can now be handed off with a description and a budget cap.
The catch is that you still need to understand what you’re buying. Token costs compound quickly across many agents. A poorly defined task can spin up more parallelism than you intended. Good evaluation benchmarks and conservative defaults are still your friends.
The trajectory is clear. AI automation is moving from “run this script” to “handle this project.” The infrastructure to support that is growing fast, and businesses that start experimenting now will have a real advantage over those who wait.
Want to explore how AI workflow automation could benefit your business? Let’s talk.