Teams of AI agents that plan, execute, and review entire projects: Gremium.ai, our own product, built by a small team in under six months

The challenge

Knowledge workers, consultants, and project managers run research-heavy projects. They spend most of their time in the same loop: break the work into tasks, research, write, track the status, and repeat. AI could take on each of these steps. A chat window is the wrong shape for that work. A single conversation cannot hold a multi-week project with dependencies, review steps, and real deliverables.

Our thesis for Gremium.ai, transfactor's own product: the correct unit of delegation is the whole project, not a single chat message. A user hands an entire project to a team of AI agents. The agents plan it, work it, and review their own output. The human decides how much oversight to keep, per project, not in a single global setting. Our internal target for this audience: a reduction of manual work by up to 80%.

The hard part is not an answer from an AI. The hard part is coordination: a breakdown of the work, the dependencies between tasks, a review of the quality, recovery of stuck work, and enough control for a person to trust the system.

What we built

Gremium.ai is a workspace. A user creates a project and hands it to a team of AI agents. A plan agent breaks the project into tasks and works out the dependencies between them. Deep research therefore happens first, and the report comes last. Execution agents then work the tasks with 20 specialized tools: web search, deep research, file access, document creation, and charts. Review agents check the quality of every result before it counts as done.

The outputs are real deliverables, not chat transcripts. A built-in report writer plans the structure of a report, fills it in chapter by chapter, adds real charts, and runs a final edit pass. The system then converts a finished document automatically into a presentation deck and a web version. A user shares a report with outside stakeholders. Those stakeholders ask questions, and the answers come only from the report itself.

All of this is visible in real time. The workspace shows the tasks on a board, a live map of the dependencies between tasks, and the current action of each agent. A built-in assistant also manages projects directly for the user.

How we built it

Every task moves through a defined series of stages: planned, in progress, review, and completed. The system therefore knows the exact state of each piece of work. Separate services handle the plan, the execution, and the review independently. A safeguard watches for stuck work and restarts it. The system runs many projects in parallel and grows with the demand.

A person stays in control through a per-project setting with three levels of agent independence. At the lower levels, an agent pauses and asks a person before a judgment call. Before that interruption, the agent can consult a more capable AI, to confirm that the question truly needs a person. At full independence, the agents run the whole project alone. Access to stored data is strictly controlled throughout, with dedicated security hardening inside the product.

What this build demonstrates

Gremium.ai is a product that works. A small team built almost all of it in under six months: roughly 72,000 lines of code. That is the point. With the correct method for the coordination of AI agent teams, a small team ships a system of this depth. Staged work, adjustable oversight, and full visibility into the agent actions are the same patterns we bring to corporate innovation engagements. In those engagements, AI must do real work under real oversight.