We are designing a four-layer system. Each layer has a single job. Together they form the substrate for an AI employee that does not reset.
This is the architecture as we have designed it. It is not fully built yet. We are publishing it openly because we believe the design itself is worth examining.
Designed so humans remain the principals at all times. The interface is built around review and oversight — not chat. Every agent output is citable, challengeable, and attributable.
A group of specialised agents, each with a defined role: reasoning, critiquing, simulating, remembering. No single agent has final authority. Claims are version-controlled and must survive challenge.
Designed to run reasoning loops asynchronously. The agents do not stop when a session ends. Hypotheses are tested, falsified, and revised continuously.
The foundation of the entire system. A queryable, auditable memory of every claim, contradiction, and open question — never wiped, never reset. This is what makes the agents employees rather than tools.
Project Laplace