Dynamic workflows move the agent plan from context into executable code

Anthropic's dynamic workflows have Claude write a JavaScript orchestration program that launches subagents, collects results and manages information flow. The documentation describes tens to hundreds of agents in one run, with a script that can be inspected, saved and run again.

Claude records coordination in a program instead of a long conversation

A lead agent first designs the approach and writes a program that divides the work. A runtime then launches agents on focused tasks and keeps intermediate results in script variables. The main context does not need every step returned to it, only the coordinated result.

At AI Engineer New York, Anthropic demonstrated a stock watchlist scanner. One phase distributed tickers to agents with isolated contexts and a later phase reviewed their findings. The company also positions the technique for codebase-wide audits, migrations and research that cross-checks sources.

Dynamic workflows were introduced on May 28, 2026, and Anthropic's primary page now marks them generally available. They are offered in the Claude Code CLI, Desktop, the VS Code extension and through the API, as well as Amazon Bedrock, Google Cloud and Microsoft Foundry. Access is tied to paid plans or API use.

A repeatable script changes the job most for engineering leads

A conventional subagent leaves the plan inside a conversation. A workflow moves loops, branches and review order into readable code. A team can rerun the same process, edit it and inspect which phase produced a result. That is more operationally useful than hoping a long prompt elicits the same behavior every time.

For engineering leads, control moves to a different layer. They do not have to approve every intermediate step, but they do have to define task boundaries, budgets and acceptance conditions. Anthropic also supports independent verifier agents, allowing review to live inside the program rather than at the end of a pile of outputs.

More agents multiply both cost and a badly framed objective

Anthropic explicitly warns that workflows use substantially more tokens than a typical Claude Code session. The conference presentation added that the model does not know its budget while running, although a developer can set a session budget. A mistaken objective can therefore fan out quickly and expensively.

Visibility does not guarantee correctness either. Subagents share a filesystem, but according to the presentation a developer cannot use the API to direct one workflow subagent to message another. Quality still depends on task decomposition, context isolation, tests and whether the verification phase uses genuinely independent criteria.

Token bills and rejected findings will reveal the real gain

A useful deployment must measure more than the number of agents launched. Relevant signals include the cost per completed migration, the share of findings rejected in human review, time saved on repeated runs and the ability to resume without losing work.

The strongest evidence will come from teams that reuse a saved workflow and maintain it with the discipline applied to other production code. If every large task ends with a one-off swarm and manual cleanup, orchestration has merely moved the queue from GitHub to the API bill.

Lilith's verdict

Anthropic has turned agent swarms into an executable timetable. If the route is wrong, a hundred passengers merely reach the wrong station faster.