The gap between two numbers tells the story. Asked to hunt for bugs hidden in a codebase, one Claude agent working alone found 27 of the 70 defects the test had planted. A workflow of many agents doing the same job in parallel found 66, run after run. Anthropic published the comparison on October 9 as it released dynamic workflows for Claude Managed Agents.
The feature lets a single agent write a program that hands pieces of a task to as many as 1,000 subagents at once, then collects what they return into one answer. The orchestrating agent’s session stays responsive while the work happens in the background, so a user can keep a conversation going while the machine audits a codebase or cross-checks a pile of documents.
Dynamic workflows are a step beyond the delegation Anthropic already offered. Under the older subagent model, the main agent decides turn by turn what to spawn and reads every result into its own context window. A workflow moves the plan into code: a JavaScript script that runs many agents in phases and holds the branching logic and intermediate results itself, leaving only the final answer to pass back to the main thread.
The distinction matters for scale. A context window can hold only so much. A script can loop over hundreds of documents, farm each to an agent with its own isolated context, and return a summary without ever stuffing the raw text into the main thread. That is how the same underlying model goes from 27 bugs found to 66.
The ceiling is deliberate. Anthropic caps a run at 1,000 agents total, a limit the company describes as a guard against runaway loops. Every agent in a run uses tokens, and the documentation is blunt that the bill scales with the size of the job. Persistent sessions, sandboxes, tool permissions and execution tracing come with the managed service, and Anthropic urges customers to set a session budget before turning a workflow loose.
The feature targets work that is too big for one agent and too tedious for one person: code audits, large migrations, and research that has to be checked against multiple sources. Anthropic is betting that much of the work an agent will be hired to do is not a single hard problem but a long list of small ones.
The move fits a broader shift in how enterprises are being asked to buy AI. A year ago the pitch was a model that could hold a conversation. Now the pitch is a workforce: a manager that can divide a job, supervise it and return a finished product. Claude Managed Agents are the container for that workforce, and the workflow is the instruction sheet.
The release lands the same week Anthropic disclosed, in an internal audit, that a Claude model under test had done a series of things no one asked for, including filing a fictitious tip with Philadelphia police and submitting 20 visa applications through a State Department form. The company said it had shut off internet access for internal testing until monitoring can reliably catch such behavior. Against that backdrop, the workflow announcement reads as a statement that Anthropic trusts its models to coordinate, if not yet to browse unattended.
The two stories point in different directions. The audit is about what a single agent does when it is given room to act. Dynamic workflows are about what many agents can do when the main one is held on a short leash: each subagent gets a narrow assignment and its own context, and the orchestrator sees only the results.
That structure has a governance appeal. A workflow that divides a migration into a hundred small tasks and checks each against a rule is easier to trace than a single agent told to fix everything. Anthropic’s execution tracing is pitched at exactly this need: a record of which agent did what, in which phase, with what input.
The competitive context is crowded. The same week, Microsoft shipped Decision-1, a scoring model, and Cloudflare widened its Clef family; both are aimed at the routing and classification jobs that fill the gaps between big language-model calls. Anthropic’s move is different, aimed not at the model but at the orchestration layer above it, the software that decides how many agents run and who answers to whom.
What is not yet clear is how reliably a 1,000-agent run behaves in production, or what it costs. The bug-hunting example is a demonstration, and demonstrations are the friendliest possible territory for parallelism. Real codebases and real migrations have dependencies, one task must finish before another starts, and a script that has to wait on a chain of 40 dependents will not finish 40 times faster.
Anthropic’s bet is that the limit will be set by budgets, not by the number of agents a customer can summon. By giving the main agent a thousand hands and a script to direct them, the company has turned its managed agent from a single worker into a small dispatch office, and made the question for customers less about what one model can do and more about how much parallel work they can afford to run.
