Give AI the Goal and the Limits, and Let It Staff the Team── The AI that has opened the work is better placed to decide headcount and roles, so write the goal, the finish line, the limits and the stop points, and delegate the arrangement

01The operator named no headcount and no roles, and left even the choice to use a team to the AI

On 11 September 2026, the operator added one permission to a request already in progress: if it helps, you may run a team of agents, and how you staff it is up to you.

A team of agents here means a setup in which one AI splits the work and hands the pieces to other AIs. We will call the AIs that take those pieces workers, and the one that splits and assigns the lead.

The sentence leaves a great deal unsaid. How many workers? Who takes which part? In what order? Should there be a team at all? None of that is written down. What is written down is the goal, already set by the earlier request, and the limits of what the AI may do.

The question is whether this is a lazy way to delegate or a pattern with reasons behind it. It is a pattern. But it stops being one the moment the limits are left out and only the staffing is handed over.

02Only the means were handed over; the goal and the limits stayed with the person

A request to an AI can be read in three layers. The goal: what you want achieved. The means: who does what. The limits: how far the AI may go.

The operator's sentence gives away only the means. The goal was fixed by the work already under way. The limits were stated, with a condition attached: a team only if needed. What the operator let go of was the arrangement of the work, nothing more.

The difference from the right-hand column matters most. Handing over only a goal is closer to walking away than to delegating. The AI's judgement has a defined territory only once the limits are written.

03The right split shows only after the work is opened, so a person who fixes it in advance often gets it wrong

Staffing is hard to decide from the outside because the right split depends on what the work contains. How many files are there? Do they depend on each other? Can they be handled side by side? Whoever has started the work finds this out first.

In a December 2024 article, Anthropic describes a pattern in which a lead model breaks a task down and hands the pieces to workers. It suits tasks where you cannot predict the subtasks in advance. The subtasks are not defined beforehand; the lead decides them after seeing the input.

Staffing also has a price. In the multi-agent research system Anthropic described in June 2025, a team of AIs beat a single AI by 90.2% on the company's internal research evaluation. It also used about 15 times as many tokens as an ordinary chat. A single agent using tools used about 4 times as many.

The same report records an early failure: the system spun up 50 workers for simple questions. A team does not make every task better. That is why the words "if it helps" in the operator's sentence carry weight. They leave the decision not to use a team with the AI as well.

04A request that delegates staffing has four parts: goal, finish line, limits and stop points

In this series, a request that delegates staffing means this: you write the goal and the condition for being done, you write what may be used and which points may not be passed without a person's approval, and then you let the AI decide headcount, roles and order.

There is a boundary to draw. What you delegate is the staffing, not the responsibility for decisions. Beyond the stop points, a person decides. OWASP, a security community, lists giving an AI more functionality, permissions or autonomy than it needs as one of the main risks of AI applications, and recommends requiring human approval before high-impact actions.

It also helps to separate this episode from others in the series. Having the AI list the items of a job before building is episode 4. Deciding yourself the order in which several requests run is episode 6. This episode covers one thing: letting the AI decide who does what inside a single job.

05In material review, the AI can staff the review by viewpoint while the verdict stays with people

Take promotional material review at a pharmaceutical company. A single product brochure has to be checked from several viewpoints. Do the efficacy claims stay within the approved indication? Do the figures match the cited papers? Is the safety information adequate? Does the English version say the same thing?

Whether to give each viewpoint its own worker, or to have one AI go through them in turn, depends on the size of the material. A few pages need only one. A large document or a revision with many changes may go faster when split. That call can be left to the AI that has opened the material.

The same pattern works for literature searches in medical affairs. Following several search directions side by side is the kind of broad research that Anthropic's report found suited to multiple agents. In marketing, when drafting for several channels, the AI can decide whether to split the work by channel.

In every case, the reason to write the limits first is the same. Company rules restrict which documents an AI may read. However many workers there turn out to be, that restriction must not change.

06Your one sentence is copied into the instructions of every worker

Now look at how a delegated request actually runs. In Claude Code, Anthropic's coding tool, each worker runs with its own context, its own instructions, its own tool access and its own permissions. The lead reads each worker's description and decides which one to send work to.

A worker does not receive the person's original request. It receives a task description written by the lead. Anthropic's report says each worker needs an objective, an output format, guidance on tools and sources, and clear task boundaries. Without them, workers duplicate each other's work and leave gaps.

This tells us something. The goal and limits you write pass through the lead and are copied into each worker's description. If your sentence is clear, that clarity reaches every worker. If it is vague, the vagueness is multiplied by the number of workers.

The limits travel the same way. In Claude Code, each worker can be restricted to certain tools or barred from writing files. If you state the limits, the lead can carry them into each worker's settings. If you do not, there is nothing to carry.

07From tomorrow: write the goal and limits, see the staffing plan first, and get a report on who handled each part

There are four steps.

1. Write the goal and the finish line. In one or two sentences, say what the work is for and what done looks like.
2. Write the limits. Which material may be read, which places may be touched, how much may be spent. Say a team may be used if needed, so that not using one stays an option.
3. Write the stop points. Before deleting, publishing or sending anything outside, the AI asks for approval.
4. Ask for the staffing plan first and a report at the end. Before starting, have the AI list how many workers it will use and who does what. When it finishes, have it report who looked at what, and what nobody looked at.

The fourth step has a reason. Seeing the plan first lets you stop an oversized setup, such as 50 workers for a small task, before it starts. The final report shows the gaps between workers. You hand over the staffing, but you keep two chances to look at it.

Widen what you delegate as the results come in. Start with work that only reads, and move to work that writes once the reports prove reliable. Anthropic's 2024 article gives the same advice: find the simplest solution possible, and add complexity only when it is needed.