/ Introduction
How it works
One request travels from you to the CEO, out to the departments and their specialists, through QA, and back to you as a sourced report.
One request, step by step
- You ask the CEO. One message in plain language, such as "Compare three suppliers of 250 g coffee bags on price, minimum order and delivery time." The CEO is the only agent you need to talk to.
- The CEO weighs it and plans. Light questions get an answer right away. For real work, the CEO opens a project, picks the departments it needs and gives each its part of the brief. For hard calls, the CEO can ask a council first (see below).
- The departments do the work. Each lead breaks its part into tasks and assigns the right specialists. Specialists search and read sources, write, calculate or draft.
- QA checks it. Every specialist's output is reviewed before it is used. Work that falls short goes back with notes, up to two revisions.
- The leads summarise. When every task is done, each lead writes the department summary, and QA checks that too.
- The CEO writes the final report. From the department summaries, the CEO writes the report with its list of sources. QA checks it once more, and the report lands in your chat.
While the work runs, you can follow it in the Office and on the project page. When the team needs a decision from you, such as the lowest price you would accept, the question waits for you in Needs you.
Project depth
When it opens a project, the CEO picks a depth that fits the request:
| Depth | Good for | Specialists per department |
|---|---|---|
| Quick | Clear, narrow questions | 1–2 |
| Standard | Most research and content requests | 2–3 |
| Deep | Big decisions that need many sources | 3–5, plus a second, independent review |
Every project has a token budget, so one request cannot run up costs unchecked. See Costs and limits.
QA review
QA is an agent that does none of the work, so its review takes no side.
- Every deliverable is reviewed against its task's criteria, including whether its claims are backed by sources the team actually read.
- Up to two revisions. QA's notes go back to the specialist. Whatever still does not pass is flagged in the report, not hidden.
- A different model. QA works best on a different model family than the specialists, for example specialists on Claude and QA on GPT. That is why we recommend connecting two model providers. With only one, QA still reviews, and the Usage page tells you.
- The team learns from QA. QA's feedback is kept as lessons for that specialist role, so the next specialist does not repeat the same mistake.
A council for hard calls
For decisions without a single right answer, the CEO can call a council of three viewpoints: a Skeptic, a Pragmatist and a Critic. They weigh the options before the CEO decides, and you can read their discussion in the chat.
You stay in control
- The team works on projects only when you ask or on a schedule you set. Background memory upkeep is set in Settings → Automation.
- Social posts an agent wants to publish always wait for your approval. Changes in outside services such as Notion wait for approval too, unless you turn that off per integration.
- Every project, task, revision and source is on record for you to trace.