Intelligence+ durability= capability.

Durable Intelligence is an AI research lab. In pursuit of an abundant future, we study AI durability for both productivity and safety.

The intelligence is ready. Durability is the other half.

To achieve abundance, we must build an aligned and durable foundation, so that intelligence creates new productivity which grows in stride with human enlightenment.

Our founding belief

Frontier models reason well enough for serious work, but they can’t be trusted in an enterprise setting, in public services, or even in managing your own life. What limits their use is the underdeveloped frontier of durability: the ability of agents to keep their goals, constraints and work valid as a project runs and changes.

We study safety as a precondition for productivity: an agent’s ability to be trusted is a product of it being safe to use, and dependable enough to be given work. We are working towards solving both by developing the model’s external “prefrontal cortex”.

In 1991, Shallice and Burgess studied three patients with injuries to their prefrontal cortex who scored well on intelligence tests. Sent on a simple errands task with its rules written on a card they carried, they regularly broke those rules. Their knowledge was intact. Executive function, the capacity to keep a goal and its constraints active over time, was not.

Long-running agents show the same gap, and we call the missing property durability.

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.

Science is the most durable system we know, and we are building agents that work the way it does. Researchers leave; their methods, records and review outlast them, so each generation builds on checked work. Agents need the same property across sessions and model generations: capability that compounds, with its constraints and the means to verify it carried forward. Agents built that way could become how science itself gets done.

Read the thesis

What we mean by durability

  1. Long-horizon reliability

    An agent can be trusted to take over projects with multiple stages of sub-tasks, reasoning, and decision making, over weeks and months.

  2. Models as co-workers

    An AI colleague learns how the project works, knows why decisions were made and acts when circumstances change.

  3. Safe delegation

    Less human intervention and hand-holding, so that human input is leveraged for creative or highly-complex strategic guidance, and standard decisions for project executions can be delegated to responsible models.

  4. Systems of intelligence

    We’re testing whether models and agents can work as one organisation, with discoveries that outlast the agents who made them. Similar to human societal learning over generational succession.

  5. Sharp to the last decision

    A model reasons as carefully late in the work as it does at the start, with the relevant evidence available for each decision, to ensure consistency of performance.

  6. Knowing what is still true

    New evidence changes the decisions that depend on it. The team stops using outdated guidance and acts on what applies now.

Where durability breaks, from the model up

A drawing of a language model at work inside an agent stack: a sealed model reads a tape of text through a fixed window, with its memory cabinet, a press that compacts the window, a guard gate on the way to the tools, a checkpoint rail, and a newer model hanging ready.

The model

  1. Next word

    A language model predicts the next word, one word at a time.

  2. The window

    It only sees a fixed-size context window. As the work grows, the oldest text falls out.

  3. Rot

    Attention compares every word with every other word, so cost grows with the square of the window. Models also get worse as inputs grow: all 18 models in Chroma’s study degraded.

  4. No memory

    Each call starts blank, and the model’s weights do not change while it works.

  5. Unenforced

    A rule is just text. Nothing makes the model follow it, and earlier instructions hold less well as conversations grow: scores fell from 90% to 65% across 15 models (Laban et al., 2025).

The agent stack

  1. The harness

    The harness runs the loop: plan, call a tool, read the result, repeat. It decides what goes into the window.

  2. Memory

    Memory systems store notes, documents and logs outside the model and retrieve what looks relevant. Everything still has to pass back through the window to be used.

  3. Context management

    When the window fills, context management compacts it into a summary, trims old turns or hands work to a sub-agent with a clean window.

  4. Guardrails

    Guardrails check actions against fixed policies and allowlists, and can ask a person for approval.

  5. Checkpoints

    Checkpoints save the session so it can be resumed or rolled back.

Where it breaks

  1. Drift

    Long runs cycle through compaction. Each cycle drops a rule or a reason, and the agent drifts from its goal. Bigger windows rot, and better summaries still lose detail.

  2. The missing rule

    Retrieval fetches text that looks similar. A rule worded differently is never loaded, and the step runs without it.

  3. Broken in plain sight

    The rule is in the window but carries little weight, and guardrails check fixed lists, not commitments made during the task. The action goes through.

  4. Stale decisions

    Checkpoints save text, not which decisions depend on which facts. When evidence changes, nothing knows what to revisit.

  5. Upgrades start over

    A newer model inherits the transcript, not the reasons, so it reverses decisions it cannot explain.

The durability gap

  1. What the stack covers

    Each part covers a piece. None keeps commitments valid over a long job.

  2. Why it matters

    That gap is the durability problem: broken rules, repeated work and people re-explaining the project. In our September 2026 paper, a leading agent framework with memory and checkpointing kept 63.9% task fidelity at day 30.

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.

Beyond the Horizon: what durable AI could make possible.

1 hour · 1 day · 1 week · 1 quarter · 1 year · 1 decade · 1 century

Durable AI is the next great leap to models that can be trusted to learn, create, discover, and work alongside us.

Dawn of Durable Intelligence

Six problems between frontier models and durable AI

The six questions that decide whether agents can be trusted with long work, and the focus of our research.

See the six research areas
  1. Action state

    Which evidence should govern the next decision, and what must change when that evidence changes?

  2. Latent control

    How can a model keep what it learns on the job through interruptions, restarts and new models?

  3. Model continuity

    What does a replacement model need from its predecessor to take over unfinished work?

  4. Collective intelligence

    Can a population develop abilities that survive its members and exceed what any member can do?

  5. Self-improvement

    Can a system use the methods it has learned to improve the AI that finds its next methods?

  6. Economics

    Which system finishes useful work for the least total effort, including the work people must repeat?

100 agents · 4 worlds · 30 simulated days

Capability that belongs to a population.

If an agent discovers a water source and later dies, can the group still use what that agent learned? The Valley is an experiment designed to find out: agents live in mortal bodies, sense only what is near them and leave records in objects. Can artificial intelligence properly form societal evolution under hostile survival conditions?

Read the protocol

Product of our research: The Nenya Harness

Request an invite Learn about Nenya

npx -y @qwntllabs/nenya setup

We founded Durable Intelligence to build the most critical pillar for a safe and abundant future: scaled productivity in the service of humans.

Learn more about what influenced our dedication to durability and alignment.

About the company
  • Lindell Cumes
    Lindell CumesCo-founder and CTO
  • Jacob Schoenberg
    Jacob SchoenbergCo-founder and CEO

Help build the key to abundance.

We want to hear from researchers and engineers who have pushed agents past a stable performance, whether that meant a months-long investigation or a team of agents that had to learn from each other. Tell us about the hardest problem you have hit and what you tried.

Continue reading →