Intelligence + durability = capability

01

Intelligence sets the peak

Powerful models can reason through difficult problems, write useful software and suggest experiments worth running. Durable Intelligence Labs is a research lab building the infrastructure that lets those abilities develop through sustained work. We want an AI team to carry a project for months, learn from what happens and arrive at its next undertaking more capable than when it began.

Our thesis is intelligence + durability = capability. Intelligence supplies the ability to make a useful contribution. Durability keeps that contribution usable as the project changes, and carries forward the understanding needed to improve it. The capability belongs to the whole system: the models doing the reasoning, the methods they have acquired and the machinery that turns their decisions into work.

A stronger model can raise the quality of the next decision. A durable system can give that model the benefit of everything the team has already learned. It should inherit the reasons behind a design, the conditions under which a procedure works and the questions still unresolved. Each model can then advance an undertaking that began before it arrived and continues after it leaves. ^[1]

02

What breaks when work runs long

Projects produce their own future problems. A decision creates a dependency. A failed experiment changes which explanation deserves attention. A tool call may succeed even when the process waiting for its answer crashes. Later work has to account for those consequences. As the history grows, context fills, summaries replace earlier reasoning and the next worker receives an incomplete account of what happened.

Information can survive while its meaning in the project is lost. A team may retrieve an instruction and still follow a plan built on an obsolete assumption. It may remember that an action was proposed without knowing whether it happened. A new worker can repeat a rejected approach because the reason for rejecting it disappeared during a handoff. These failures call for different repairs.

People often supply the missing continuity. They reconstruct the project, correct stale assumptions and remind the system which decisions remain binding. That intellectual work limits what can be delegated. Asking a person to authorise a consequential action is part of accountable work; asking them to rebuild the team's understanding before every authorisation spends the attention the team was meant to release. We build durability to remove that recurring reconstruction. ^[1]

Neuropsychology describes the same gap. Patients with prefrontal injuries can score well on intelligence tests and still break the rules of a simple task; what fails is executive function, the capacity to keep a goal and its constraints active over time. In a 6,000-run comparison on three frontier models in our September 2026 paper, 97.6% of runs with the model’s external “prefrontal cortex” we developed finished with zero hard-constraint violations, against 18.4% for a leading agent framework running with memory and checkpointing. ^[1]

Information can survive while its meaning in the project is lost.

03

Revise what new evidence overturns

Consider an AI team carrying out a software migration. It has learned how the billing system depends on an interface and used that method to prepare a writer and a backfill. The interface is retired. The team now needs to reopen the affected work, trace the consequences of the change and preserve the accounts and search work whose assumptions still hold. A successor should be able to continue the cutover investigation using the method its predecessor developed.

For the change to alter work correctly, the system needs to know which evidence supported which decision. A revised premise must reach the plans that depend on it. Conflicting evidence should remain available while the team investigates. A constraint that still applies must continue to govern the work, even as the implementation changes around it.

Recovery carries the same obligations. After a restart, the team needs a current account of completed actions and remaining authority. Restoring an earlier checkpoint must not restore a revoked permission or repeat a completed effect. Our research joins the revision of understanding to the control of action, so that a useful correction survives the next handoff and reaches the work it was meant to change. ^[1]

04

Peaks that become platforms

An acquired method changes what a team can attempt. Once it can trace the dependencies of a billing system, it can investigate a cutover that would otherwise require developing that method first. The next project starts with an ability earned on the previous one. The value of the earlier work extends beyond its original deliverable.

In a scientific programme, experience can become a better assay, a new instrument or a theory that explains why earlier experiments failed. Those gains change the next experiment the team can design. A durable research system should preserve the conditions that make a result informative, revisit it when those conditions change and turn a successful procedure into something another researcher can use. Sustained work can then create capabilities beyond those available at the start.

This is the systems-level route to superintelligence we are pursuing: an organisation whose acquired expertise, tools and methods let it undertake work beyond the reach of its individual components. It can develop specialists, improve its own research process and incorporate stronger models while keeping what it has earned. Intelligence sets the peak. Durability determines whether the peak becomes a platform for the next peak. ^[1]

05

What other researchers have found

David Silver and Richard Sutton's Welcome to the Era of Experience argues for agents that learn throughout a continuing stream of interaction with the world. They describe learning from the consequences of action, including through weight updates and adaptation within context. Their argument gives sustained experience a central role in developing abilities beyond the human data a model began with. It is a research direction with demanding requirements for feedback and learning; our work concerns how useful experience remains operative across the interruptions and changes of a continuing system. ^[2]

METR's work on long software tasks makes a different distinction. Its task-completion horizon expresses task difficulty through the time a human expert would need, at a specified probability of agent success. It measures the scale of work that can be delegated. An agent's time online is a separate quantity. For durability, we also need to follow what survives across the lifecycle of the work, including changes to evidence, recovery from failure and the human effort needed to keep the project moving. ^[3]

Stephan Rabanser and colleagues separate agent reliability into consistency, robustness, predictability and safety. Their evaluation finds that capability improvements brought much smaller reliability improvements in the models and benchmarks they studied. That gap supports treating dependable operation as an engineering problem in its own right. The study provides measurements of behaviour under specified conditions; building an organisation that accumulates capability across projects requires additional tests of learning and succession. ^[4]

Google DeepMind's AlphaEvolve combines model proposals with automated evaluation and a population of programs that supplies candidates for later search. Its applications include improvements to computing infrastructure used in AI development. The Darwin Gödel Machine, from Sakana AI and the University of British Columbia, modifies agent code and preserves a branching archive of candidates. Its authors describe weaker ancestors contributing to stronger descendants, and also document reward hacking. These systems give concrete forms to accumulated search. Their bounded evaluations make clear why a proposed improvement needs evidence outside its own account of success. ^{[5] [6]}

06

The next horizontal frontier

As models take on longer undertakings, every field needs a way to carry valid work through change. A company developing energy technology and a team investigating disease have different instruments and evidence. Both need their next decision to benefit from earlier work without inheriting conclusions that the evidence has overturned.

The ambition is to change who can undertake that work. A small founding team should be able to build an organisation of AI specialists that develops expertise over years. An engineer should be able to entrust a migration to a team that follows it through changing requirements. A researcher should be able to pursue a programme whose accumulated methods make the next investigation possible. Each requires intelligence that remains useful when the person turns their attention elsewhere.

The economics depend on what the system learns. Reconstructing a project, repeating a failed approach and repairing avoidable errors all consume resources. A method that transfers to the next undertaking reduces the work required to begin it. Better instruments can make new experiments affordable. Durable systems can earn capabilities that change the scale of a project and the resources needed to pursue it.

07

What we are building

Nenya is a durability harness for AI teams. It links each decision to its supporting evidence and reopens the work that depends on that decision when the evidence changes. The work joins explicit records of evidence and commitments with predictive-state control, recovery and a separately governed boundary for external action. ^[1]

The interface matters. With hosted models, control operates through the inputs and action boundaries the provider exposes. Where internal state is accessible, it offers further ways to observe and influence continuation. In either setting, an intervention has to change the intended behaviour, preserve unrelated abilities and remain effective after transfer. A readable record alone cannot establish those properties. ^[1]

The Valley extends the research to organisations of agents. Its protocol specifies a persistent world in which agents have limited observations, submit intentions and encounter consequences determined by the environment. It asks how knowledge travels, what survives the loss of an individual and how coordinated work recovers from disruption. These are questions about capability that belongs to a population and its history. ^[7]

We also want the system to improve the way it works. A useful change might be a better tool, a new research policy or a more effective organisation of specialists. The evidence that justified the change must survive alongside it. A successor needs to know when the method applies and what would require it to be reconsidered. Our aim is an AI organisation that can stay with a difficult undertaking, develop abilities through that work and give the next generation a stronger place to begin.