Who Controls The Flow Is The Whole Story — And The Middle Column Has The Wrong Boss

This card has the best subtitle we have read on a Claude Code topic. Three columns — sub-agent, agent team, dynamic workflow — under one line: “Same workers, different boss. Who controls the flow is the whole story.”

That is the correct question, and it is the one Anthropic’s own documentation uses to separate these three. The workflows page has a comparison table whose second row is literally headed “Who decides what runs next”. So the card and the docs agree on the axis.

They do not agree on the answer in the middle column.

Diagram: who decides what runs next

Two out of three is a good score, and the two it gets right it gets right in the right words. But the middle one is not a detail. The whole value of the card is that it tells you which tool to reach for, and the picture it draws for a team — IN → Strategist → Designer → QA Review → Publisher → OUT, a four-step relay with numbered batons — is not what an agent team is.

A team is not a relay. It is a room.

From the agent teams documentation, in its own words:

“Teammates work independently, each in its own context window, and communicate directly with each other. You can also talk to any teammate directly without going through the lead.”

“Agent teams are most effective for tasks where parallel exploration adds real value.”

The examples the docs give are a parallel code review with three reviewers applying different lenses, and five teammates testing competing hypotheses who are explicitly told to “talk to each other to try to disprove each other’s theories, like a scientific debate”. Not a chain. A room full of people arguing.

And then there is the sentence that turns the card’s middle column inside out. The card says a team wins at “repeatable pipelines where the same specialist roles run every time”. The documentation says:

“Agent teams add coordination overhead and use significantly more tokens than a single session. They work best when teammates can operate independently. For sequential tasks, same-file edits, or work with many dependencies, a single session or subagents are more effective.”

A fixed pipeline is a sequential task with many dependencies. The card’s use case for column two is the case the docs name for columns one and three. If you genuinely want Strategist then Designer then QA then Publisher, every time, the mechanism for that is a workflow script — column three, where the order lives in code and is repeatable by construction.

And on a default install, nothing happens

There is a second thing about column two that the card does not mention, and it is the one that will cost you ten minutes of confusion. The agent teams page opens with a warning box:

“Agent teams are experimental and disabled by default. Enable them by setting CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1… Without that variable, no team is set up at session start, no team directories are written, and Claude does not spawn or propose teammates.”

So the card’s create agent team does nothing until you set an environment variable it does not name. And the other half of that line — “one .md file per role” — describes a different feature. The team’s own config file is generated, not authored:

“The team config holds runtime state such as session IDs and tmux pane IDs, so don’t edit it by hand or pre-author it: your changes are overwritten on the next state update. To define reusable teammate roles, use subagent definitions instead.”

Subagent definitions are one Markdown file per role, in .claude/agents/. The card has taken a true fact about one feature and printed it under another. The practical version is: write your roles as subagent definitions, then name one when you ask for a teammate.

Where does “up to 100 agents” come from?

The workflow panel carries a badge reading “up to 100 agents”. We went looking for that number and could not find it. The documented figures are different, and more useful, because they tell you where the friction actually is.

The card’s own footer anticipates exactly this, incidentally: “Feature names, commands and limits (such as the number of agents) change often. Check Claude’s official docs for what is available to you.” It is the right disclaimer, printed under the wrong number.

The corrected version of the card

None of this means the card should be thrown away. Fix one column and it becomes genuinely useful:

• Sub-agent — Claude decides, turn by turn. For quick delegation where only the result matters. As printed.
• Agent team — peers who talk to each other, coordinated by a lead that decides turn by turn. For parallel exploration: three reviewers with different lenses, five investigators arguing over a bug. Not for pipelines. Experimental, off by default.
• Dynamic workflow — your script decides. For scale, and for the repeatable ordered pipeline the card drew in the wrong box. As printed.

Read that way, the subtitle survives intact: the workers really are the same, and the boss really is the whole story. There are just three bosses — Claude, a lead agent, and your script — and the middle one is a chairperson, not a foreman.

Four honest caveats

1. We did not run a team. Everything above is the card checked against the agent teams and dynamic workflows documentation. We have not measured whether a team behaves more sequentially in practice than the docs describe.
2. Experimental features move fastest. Agent teams carry an explicit experimental warning and a list of known limitations. A card written a month from now could reasonably say something different, and the environment variable may stop being necessary.
3. “Up to 100 agents” may not be wrong so much as unsourced. We could not find it in the documentation; that is not the same as proving no surface has it. We report the five numbers that are documented.
4. The card is not describing fictional features. Sub-agents, agent teams and dynamic workflows all exist, all do roughly what the card says in two columns out of three, and the framing is better than most. The objection is to one column’s controller and one use case, not to the premise.

The one-line version

The card asks the right question — who holds the plan — and answers it correctly for sub-agents and workflows; but an agent team is not a roster you walk in order, it is a set of peers that message each other while a lead decides turn by turn, and the fixed pipeline the card sells it for is the exact case the documentation tells you to solve with the other two columns.

If you want an ordered chain that runs the same way every time, that is column three. Column two is for arguments.