A job title is not enough to define an agent. ‘Researcher’, ‘writer’ and ‘reviewer’ are useful only when each has a distinct assignment and recognizable output. This guide proposes a way to design small teams. The example roles are illustrative, not a list of agents bundled with ConductLoop.

Define a role by its deliverable

Start with the document or result you need. A researcher could return a findings table with links, dates and uncertainties. A writer could turn that table into a draft for a defined audience. A reviewer could identify unsupported statements and deviations from the brief. Those deliverables distinguish roles more clearly than an instruction to act as a world-class expert. Concrete scope also makes it easier to see when one role is duplicating another’s work.

Specify inputs and dependencies

For each role, write down the material it receives and whose output it depends on. If a writer needs a researcher’s findings, starting both before agreeing on sources may create duplicated work. Reviewing existing documentation and collecting existing user questions, however, can be independent tasks. Making that distinction helps a coordinator order the work. Every task does not need to pass through every role; some can finish with one deliverable.

Begin with a small team

An initial arrangement could use a CEO coordinator and two agents with different responsibilities. Add another role when you can explain which specific failure it prevents or which new output it produces. If coordination spends more effort reconciling responses than advancing the goal, reconsider the division. Keeping a team small makes its behavior easier to observe and its results easier to compare before introducing more dependencies.

A reusable role card

Role: documentation reviewer.
Receives: brief, draft and selected sources.
Returns: prioritized issues with excerpts and proposed fixes.
Criteria: accuracy, clarity and consistency with published material.
Out of scope: changing the goal or publishing the document.
Escalation: flag contradictions requiring a human decision.

You can use this card in a conversation with one model. It does not require a multi-agent system to be helpful: it makes expectations for the review explicit.

Avoid circular approval

Agreement between two agents does not establish that a claim is correct. Provide criteria and evidence for a review, and ask for concrete errors. A review that only says ‘looks good’ conveys little information. You can assign a role to compare the output with the brief, but examine whether it finds useful issues. Do not add another review solely to obtain an additional approval.

How this relates to ConductLoop

Boardroom already organizes project agents and tasks. The hierarchical integration we are preparing will extend the idea of building a team and directing it through a CEO. This guide does not confirm a role catalogue, team-size limits or per-agent model configuration. Start by documenting role cards in your notes and observing which ones help with actual tasks. The briefing guide provides a companion template for the assignment such a team would receive.