Research
Our research studies durability: how AI systems keep the rules they are given across long work, and how what they learn can outlast any one agent or model.
Papers
Measuring what durability adds to frontier models
We held the base model fixed and measured what durability alone changes over month-long professional work.
Can what an agent learns outlive it?
A study protocol for four worlds of a hundred agents each, to test whether knowledge and skill survive the agents that hold them.
Measuring when agents forget the rules they were given
Experiments on frontier models from Anthropic and Google show that an agent loses a rule at one moment: when it leaves the context window.
A technical summary of Nenya
The July results in five pages: the problems long-running agents meet, and what changes when their commitments are kept in force.
Why agents get pulled back from production
Agents are often withdrawn after they reach customers. We trace the cause to commitments that slip over the length of real work.
Why capability isn’t what limits long-running agents
Frontier models have capability to spare. What long, regulated work lacks is fidelity to the rules a model knows, held over time.
Two ways agents break the rules they were given
Agents fail rules in two different ways: they lose them, or they keep them and act against them. Each needs its own remedy.
Holding an agent to a month of growing obligations
A thirty-day model of five professional fields asks whether an agent can carry a growing load of obligations without losing coherence.
Holding a closed frontier model to hard rules
A frontier model can be held to hard rules from the outside, with formal guarantees, as long as checking them is compulsory.
Research areas
8 papers
Action state
Keeping each decision tied to the evidence and the rules that govern it.
4 papers
Latent control
Keeping what a model learns on the job through restarts and new models.
8 papers
Model continuity
Handing an ongoing project to a new model without starting over.
2 papers
Collective intelligence
Abilities that belong to a population and outlive its members.
3 papers
Self-improvement
Using what a system learns to improve the system that learns next.
7 papers
Economics
What it costs, in compute and in people's time, to finish long work.