The distance is not infinite. The gap between intelligence and durability is finite, measurable, and closable.
In three years, the world will be run by superintelligence. ASI is coming. Humanity will inherit an existence free of disease, labour, and want.
Yet, ahead we face potential collapse of capital markets. The Mag Seven will crumble, and the global political climate will grow even more unstable. In front of us lies a chasm where agents never make it to economic return, AI becomes too expensive, and agents produce little to no upside for the vendor, user, or society at scale.
Intelligence is not enough.
The skeptics are right that there’s trouble ahead. The optimists are right about what ASI will deliver. The variable that sits between the two camps is durability, and it is measurable.
We have spent the last three years obsessing over how intelligent models can become. The more interesting question is how long they can maintain it. Capability is not the model’s raw power alone. It is the difficulty of the tasks it can accomplish, the problems it can solve, and its ability to sustain that level of performance long enough to complete the work.
A model can be superhuman at the beginning of a task and still be useless at the end of it, not because the models aren’t ready, but because the system is unable to hold the conditions required to perform and to be trusted.
Even if we believe that Fable/Mythos, Kimi, or Astra (Doug) can run unaided for twenty hours, the critical question is: what happens when superintelligence hits a full context window? What happens when it compacts and attempts to resume running a country from a short summary of previous context?
True capability then must be a composite of the intelligence of the model, its context, memory, tools, environment, ability to verify itself, and longevity. Capability is the total system, and the total system is only as capable as the point at which it begins to degrade. The discussion around reliability has spent too much focus measuring what a model can do in a moment, while almost every case for delivering ROI requires it to operate across thousands or millions of moments.
We are testing for the wrong problem.
The Bubble Doesn’t Require a Plateau
The bubble camp does not need a plateau in model improvement to be right. It only needs AI to be less durable than the economic assumptions being made around its subsidization. If an agent requires a human to reconstruct its state, correct its drift, remind it of decisions, explain which parts of the world have changed, and double-check its work, then the risks of human error, as well as the burdens of human micromanagement, remain very real.
Agents are not yet autonomous workers; they are intelligent and untrustworthy tools, used by people who must serve as the memory, management, and verification layer of the agent without misstepping.
This can still produce value, and a great deal of it, but it holds its own risks and is far from the autonomous labour force currently being priced into the market. The current approach will not produce the margins, scale, or speed that the historic level of investment into AI has underwritten.
The degradation is also not obvious. An agent begins with a clean context and a complete representation of the problem. But every action creates more state, every tool introduces more information, every failure creates another branch, and every decision becomes something the agent must either preserve or later reconstruct. Eventually the context is full, the history is compacted, and what returns is a compressed interpretation of the agent rather than the agent itself. Reasons become conclusions. Uncertainty becomes fact. Abandoned paths return as active assumptions. The current state risks that the smallest error, once carried through enough compactions, grows to the size of a corporate lawsuit, and soon this same flaw becomes an existential consideration. An invisible slip in an early compaction becomes part of the reality the agent believes it is operating in and compounds over its task horizon. The agent can still write, reason, use its tools and confidently explain what it is doing, which makes the degradation even more dangerous. Because a genius agent does not announce the moment it starts functioning unreliably.
The Next Frontier Is Durability
Ultra-long-horizon performance must be the next frontier.
Durability is the capacity of an agent to hold its intended function, its judgment, its rules, its performance, through sustained operation. The word means what it means for a bridge or an engine: not whether a system survives a crash, but whether it survives use.
We need agents with the ability to work far beyond hour twenty and still retain the same judgement, purpose, and understanding as hour one. All while correctly incorporating everything that happened in between. Today, that is far from true, and it is clear to anyone who’s committed to a pilot of these early prototypes we’re calling companies.
It must be able to do this for a week, a month, a year, through model changes, tool failures, new information, and the consequences of its own previous actions.
Some of this is improving. Memory and retrieval systems are becoming more sophisticated, models are getting better at verification and reconstructing lost state, and context windows will keep growing. But a larger window only moves the boundary, and a better summary is still a decision about what information to remember and what is forgotten. A truly general autonomous worker is not achievable while it continually erases its own learnings.
Some of what we call durability today will simply become intelligence, and it should. As models become more intelligent, they will appear to be more durable as well. However, the full durability of an autonomous system cannot rely strictly on that model’s intelligence. These systems need the ability to replace the underlying model as better models become available without losing the state and progress they have already accumulated. That requires the system’s durable state to persist independently of any one model.
Scaling laws ask how much more compute is required to create the next level of intelligence.Durability asks how much of the intelligence we have already paid to create is currently being thrown away.
The frontier labs are building beautiful minds. But agent infrastructure must determine how those minds can enter an existing world, understand its state, act inside it, leave evidence behind, and pass the work to whatever mind comes next.
Every Agent Becomes a Harness
The smartest model in the room is about to matter less than the memory it keeps, the rules it runs on, and the other minds it works with. If that's true, the next leap in what AI can do will come from the system, not the model.
As every agent becomes a harness, it will increasingly be defined by everything surrounding the model rather than by the model call itself: its tools, permissions, objectives, history, and relationship to other agents. As specialized models become superintelligent inside specific domains, the next arms race will not simply be who has the smartest model, but who can maintain the best system of models without losing coherence between them. One model may plan the work, another will solve it, another verify it, and another operate what they have created, but without durability the whole becomes less capable than the sum of its parts.
We should continue scaling the transformer, and searching for whatever comes after it, but it would be reckless to make the autonomous future dependent on the arrival of an architecture that does not yet exist.
If the models we have can already reach extraordinary peaks, the parallel route is to build the infrastructure capable of holding them there.
And this raises the larger question.
Is ASI more likely achieved through the stabilization of today’s peak performance than throughthe pursuit of peak performance alone?
This is not to say that today’s models have achieved ASI, or that memory and orchestration can create intelligence that is not there. But we are defining superintelligence too narrowly when we imagine it as a property that must exist entirely inside one model. If a system can maintain the best available reasoning, route problems to specialized models, preserve their discoveries, verify their outputs, recover from failures, and compound what they learn without returning to zero at every context boundary, then the system can become more capable than any of the individual models from which it is made.
The most intelligent thing in the world may first be an organization of models held together by a durable state, rather than a single omniscient mind.
Establishing the Durability Layer
Labs build durability for their own ecosystem, but their incentives leave a gap in portability and neutrality.
If autonomous systems are to survive the swapping of the models that power them, we need their state to be portable across models, inspectable outside of the model, and verifiable by parties who do not have to trust the vendor that generated it. If durability is to become a property of autonomous systems rather than a feature of one vendor’s models, then the research, benchmarks and infrastructure around it cannot begin and end inside that vendor. This will not be trivial. What one model understands about the world may not transfer cleanly to another.
What can survive that transition, what cannot, and what a successor intelligence needs in order to inherit work without corrupting it are themselves part of the research problem.
The benchmarks must also be neutral, because the question is no longer whether a model completed a task inside a controlled evaluation, but how its capability changed through time, where it degraded, what it forgot, what it falsely preserved, and whether another model could inherit the work without corrupting it. No such benchmark exists.
That instrument, a standard measure of agent durability at ultra-long-horizons, is where our work begins. The research will be published, the instruments open, and the results reproducible.
Even a context window large enough to hold an agent’s complete history would not tell it which facts remain true, which decisions were wrong, or whether its understanding of the objective has drifted from the objective itself. The problem is the preservation and governance of state beyond what the model can see.
As a result, the durability layer requires its own research: the measurement of capability decay, the separation of memory from belief, the preservation of evidence through compaction, the transfer of state between models, the detection of drift, the recovery of systems after failure, and harnesses that can replace the mind inside them without replacing the agent itself.
It also requires a definition of continuity that does not simply preserve everything, because an agent that remembers every error forever is no more durable than one that forgets everything. The task is to preserve identity, intent, and earned knowledge, while allowing it to change its mind when the world proves it wrong. Without this layer, we may still achieve the raw capability the market is promising, while failing to achieve the autonomy required to turn it into economic value.
But the inverse is far more interesting.
If intelligence can become durable, an autonomous system no longer has to be defined by the lifespan of a context window, a model, or even the architecture on which it began. It can carry work forward, preserve discoveries made by one mind and make them available to the next, change the model underneath it without losing the world it has already understood, accumulate evidence, recover from failure, revise its beliefs, and continue work whose horizon is measured not in hours, but in years.
That is a very different conception of an agent, and it opens a very different frontier.
The models can do the work. We’re trying to make sure they can be trusted to keep doing it for days, months, and years to come.
QWNTL Labs exists to make durability an independent field of research, to measure where intelligence degrades, and to build the neutral infrastructure through which intelligence survives its model, its context and eventually the people who first deployed it.
– Lindell & Jacob