We're not just interested in bringing AI into organizations. We're investigating how organizations themselves must evolve to become native environments for intelligence.
Our product work begins with workflows, requirements, examples, human corrections, and outcome evidence. We want to preserve the knowledge needed to repeat useful work and improve it.
The immediate questions are practical. What context matters? Who reviews a change? How do we show uncertainty and keep outdated information from becoming assumed knowledge?
A company that introduces a new tool should be able to build on what its people have learned. We’re investigating how reviewed workflows, examples, and requirements can prepare successive AI systems for that work.
Candidate skills are one possible output. Each new agent still needs to demonstrate competence. Company-specific evaluations and controlled environments could let teams test representative tasks before deployment.
We would compare results with ordinary documentation and starting from scratch. The measures include task quality, setup effort, and the cost of keeping knowledge current.
What should an organization look like computationally if it is to keep benefiting from changing architectures of intelligence?
Our longer-term research explores persistent memory, executable environments, simulation, learned representations, and state prediction. World models, model-specific adapters, and latent communication may offer new ways to interact with organizational knowledge.
These are research directions. Their usefulness depends on the architecture, access to model internals, and evidence from real work. Simulation must be tested for fidelity. Latent interfaces require model-specific investigation.
Established infrastructure such as Palantir’s Ontology already connects operations and simulation. We intend to build alongside existing systems and explore how organizational representations can evolve with the intelligence that uses them.
These external sources inform our direction. They describe wider findings, not our product results.
Reports uneven business integration of AI and constraints on time, cost, and skills. The digitally engaged SME sample is not representative of national populations.
Connects usage insights to workflow improvements and measured outcomes.
Describes business-specific evaluations, expert corrections, and continuing maintenance.
A simulated software-company environment for evaluating agents on workplace tasks. It motivates our question of how connected environments reveal failures that isolated tasks miss.
Evaluates tool-using agents interacting with users under domain policies, including consistency across trials. It informs our focus on outcomes and repeatability under business requirements.
Introduces LatentMAS and explores collaboration through latent representations. It raises questions about machine-native interfaces; it does not establish a portable organizational representation across arbitrary models.