A Claude workflow found 66 of 70 planted bugs. Scope the job before you run one

On October 9, Anthropic’s Claude developers account said dynamic workflows are in public beta. A lead agent writes a plan that runs across many agents in phases and combines the results at the end. How a one-person agency maps that kind of work, keeps a review, and prices the result is in how to run a one-person agency with AI agents.

What shipped

The name in that post is Claude Managed Agents dynamic workflows. Anthropic describes it as a new type of multi-agent orchestration.

The figures below are company-reported.

A run can orchestrate up to 1,000 agents. Anthropic’s developer account says you configure an agent for this, then ask Claude to run a workflow. Claude writes the plan. The getting-started path named in the thread is Claude Code, or the templates on Anthropic’s agent quickstart.

In a test, Anthropic planted 70 bugs in a 116k-line codebase. Across three runs, a single agent found 14, then 15, then 27. A workflow found 66 bugs in each of its three runs.

The same thread says these workflows can use a lot of tokens. Anthropic’s suggestion is to start with a scoped task, then incrementally increase the complexity from there.

One test, on a codebase

The published comparison is a bug hunt in one codebase. The 66-of-70 figure belongs to that test. Anthropic did not publish a token total for the three workflow runs.

On the runs Anthropic reported, the workflow found 66 of the 70 planted bugs each time, so 4 stayed unfound. The highest of the three single-agent runs was 27.

The shape is what transfers to agency work. A lead splits a list. Separate passes take the items. One step combines them. The figure of 1,000 agents is the ceiling on a run. A list of 12 items needs 12 passes and a combine step.

On a client job, the token bill sits against the fee. A task you can describe before the run is what keeps that bill visible. Anthropic’s own line is to start with a scoped task, then incrementally increase the complexity.

Four lines before you run one

Use one agent when the job is a single piece a person can read in one sitting. One draft. One page of research notes. One set of ad variants.

Use a workflow when the job is a list of separate items, and skipping an item leaves the job incomplete. A set of competitor pages. The claims in a deck. The pages on a site you are about to send out.

Write these four lines first.

1. The question. One sentence the job has to answer.
2. The list. Every item, named, with a count.
3. Done for one item. What a single pass returns for one line on that list.
4. The stop. The most that run may spend, or how many phases you will allow. That number comes from the fee on the job in front of you.

A sketch, with no result attached: a pitch needs a claim check on 12 URLs. One agent can read the homepage and write a memo. A workflow can take one URL per pass and then combine the 12 notes. That is 12 checks and one combine step. If those 12 URLs are not written down, the task is still too open for the start Anthropic describes.

A person reads the combined note before a client sees it. The lead agent writes the plan that splits the list. The review stays with the person, which is the step how to run a one-person agency with AI agents already asks you to keep.

What to do this week

1. Pick one job you already sell that breaks into a list: a research sweep, a QA pass, or a competitive set.
2. Write the four lines on one page. Count the items. Set a stop that fits that job’s fee.
3. If you try a workflow, try it on notes and files from a past job, with that list already written down.
4. Read the combined result yourself. Use a workflow on that kind of job when the extra findings are worth the tokens against the fee.
5. Leave every other job on a single agent until its list fits on the same page.

The method in the guide stays the same: map the service, keep a review, and price the result. The change this week is a lead agent that can write the phases, up to a company-reported 1,000 per run, with a suggestion from Anthropic to start with a scoped task.