Skip to content
← Home

Research & engineering

What makes an agent dependable?

Our work on AGI-1 asks how software can preserve context, coordinate specialists, and produce evidence that a task was completed.

AGI-1 overview
Orange and blue optical strands converging in a loop
Optical study for the AGI-1 program.

Memory and identity

The architecture keeps an agent’s identity separate from a particular model. Context, permissions, and project history should survive a model change.

Generalists and specialists

A generalist coordinates a mission. Specialists contribute domain knowledge. The system must reconcile their outputs against the task the user actually authorized.

Verification

A completed task needs evidence. Evaluation focuses on observable outcomes, recovery from failure, and the cost of reaching a verified result.