Build a founder AI agent team with clear roles and handoffs

Set up a small research, drafting and review workflow with role prompts, a handoff template, tool boundaries, and a worked founder example.

A useful AI team has a shared goal, separate responsibilities, and a reliable way to pass work forward. Giving five chats job titles does not make them coordinated. The practical unit is a handoff: what the next role receives, what it must return, and how someone checks it.

Start with the fictional follow-up pilot from the previous guide. The task is to turn supplied interview notes and approved scope into a one-page brief. This tutorial works as three manual passes in an assistant. Native subagents or orchestration are an optional later implementation, not something every chat interface supports.

STEP 01

Write the shared brief before assigning roles

State the finished artifact: “A one-page pilot brief based only on source notes I01–I03 and approved scope S01.” Define what the brief must include: buyer, observed problem, deliverables, exclusions, acceptance checks, and open questions. Name the customer decision it should support.

Keep facts and decisions in a shared source document. Label each source with an ID and date. When the scope changes, update that document before running another pass. Otherwise, the researcher, drafter and reviewer may be working from three different versions.

Set a practical boundary, such as one draft and one revision for the first experiment. If crucial evidence is missing, the correct output is a focused question, not more invented detail.

STEP 02

Give each role a distinct job

Use a researcher, a builder and a reviewer. The researcher extracts evidence and uncertainty. The builder creates the brief from the reviewed evidence and approved scope. The reviewer compares the draft with both sources and the finish line. You coordinate the sequence and make the business decision.

Write role instructions around deliverables. “Act like a brilliant marketer” is weaker than “Return a table of claims, source IDs, counterexamples and missing evidence.” Give the reviewer the original inputs so it can catch omissions rather than simply judge whether the draft sounds polished.

For a simple task, these can be named stages in one chat. Use separate sessions or native subagents when the environment supports them and the separation improves the work. An AI reviewer can make the same mistake as the drafter, so its approval is not final verification.

STEP 03

Connect only the tools needed for the job

A prompt describes work; a connector gives an application access to tools or information; a skill packages a repeatable procedure. They solve different problems. A role label cannot create access to your CRM, email or repository.

For the pilot brief, start with sanitized documents and no external write access. If later work needs current records, choose an available connection with appropriate access and check the exact account, source, and action it exposes. Keep account secrets out of prompts and files.

List actions that require a human decision, such as approving public claims, sending a proposal, changing a customer record, or publishing a page. Review the concrete result before that action. Authorization should be explicit about the work being done, rather than inferred from the role title.

STEP 04

Pass a complete handoff and handle rejection

Every handoff should include the task, current artifact, source IDs, decisions already approved, unresolved questions, and what the next role must check. The next person or agent should be able to act without reconstructing the previous conversation.

In our example, the researcher identifies a follow-up issue in I01 and counterexamples in I02 and I03. The builder uses that narrow evidence instead of claiming all consultants have the problem. If the reviewer spots an unsupported revenue claim, it returns the exact sentence, the missing support, and a requested correction.

Keep one owner for the draft so two roles do not overwrite it at the same time. Independent research tasks can run in parallel later; editing the same artifact needs an explicit merge step. After a bounded revision, ask the human coordinator to resolve any remaining disagreement.

STEP 05

Run a small evaluation before expanding

Try three cases: complete inputs, a missing scope item, and conflicting interview evidence. Define success before running them. A useful workflow should produce a traceable brief for the first case, ask a focused question for the second, and preserve the disagreement in the third.

Record total elapsed time, usage cost where visible, review effort, and the errors found. Compare with doing the same task in one well-briefed assistant. Extra roles are worth keeping only if the quality or coordination benefit earns their overhead.

Repeat after meaningful changes to prompts, tools, sources, or models. Save a successful procedure with a reusable skill. Keep evidence of what was actually tested; a role diagram is not a working integration or proof that an agent completed the task.

Help me run a three-stage founder workflow: researcher, builder, reviewer. First confirm the shared brief, source IDs, acceptance checks, permitted tools, and actions that need my decision. Run the stages in order; do not imply parallel agents are running unless the environment actually supports and uses them. Pass the artifact, sources, approved decisions and unresolved questions between stages. Give the reviewer the original inputs. Stop after one revision if a material disagreement remains. Finish with the artifact, checks actually performed, remaining gaps and the next human decision.