Everything Is a Harness: When Intelligence Begins to Create Itself

Recently, while reading and using agents in my research, I have kept returning to one question:

As models become increasingly intelligent, will the principal bottleneck in AI development still be the model itself?

This is not a mature theory. It is simply an attempt to connect several observations—about models, harnesses, and agents, but also about how humans create tools, accumulate knowledge, and eventually build civilization.

For the past few years, our attention has largely been directed toward making models smarter: scaling them up, improving their data, and strengthening their reasoning.

Today, however, the central tension in AI development appears to be shifting.

The question is no longer only how to create a more intelligent model. It is how to turn the intelligence a model already possesses into action in the real world.

A wide gap remains between model capability and practical application. Everything that helps close this gap can be understood as a Harness.

1. From Model Intelligence to Effective Intelligence

In its earlier, narrower sense, a Harness referred to the scaffolding around a model: tool use, context management, memory, and the agent loop.

Today, the Harness has evolved into complete systems such as Codex and Claude Code.

Such a system does more than invoke a model. It determines what the model can observe and do, how it understands a task, how it receives feedback, and what counts as completion.

A Harness is not only the transmission system of capability. It is also the constitution of the task world.

On one side, it transmits the model’s latent capabilities into reality. On the other, it defines the world the model inhabits: its observation space, action space, permission boundaries, and criteria for success.

Meaningful intelligence, therefore, should not be measured by the model alone.

We might describe “effective intelligence” as a four-part function:

A powerful model placed inside the wrong Harness may be nearly unable to complete a task. A less capable model operating in an environment with clear feedback, verification, and recovery may be remarkably reliable.

The object that matters in future evaluation is no longer the Model alone, but:

Model × Harness × Environment × Goals

2. How Agents Grow Their Own Harnesses

I think the development of Harnesses is moving through three stages.

In the first stage, humans write Harnesses for models.

We provide tools, memory, workflows, and verification mechanisms to compensate for the model’s weaknesses in long-horizon execution.

In the second stage, agents begin to generate and search for Harnesses.

When an agent encounters a new task, it no longer merely uses existing tools. It can identify a missing capability, generate a script, skill, test, or data interface, run that Harness in the environment, revise it in response to failure, and eventually select a more effective design.

The development work being done with agents across many industries is already an early form of this stage.

We are no longer asking AI only to complete a task. We are asking it to construct a local world in which the task can be completed.

The third stage is the co-evolution of Model and Harness.

Harnesses produce new trajectories and experience. That experience enters the next generation of models. The new models, in turn, become capable of generating more sophisticated Harnesses.

The system enters a recursive loop:

Model → Harness → Experience → New Model → New Harness

Tool use, context management, and workflows that humans design by hand today may eventually be absorbed into model parameters.

But the Harness will not disappear.

Once an old Harness has been internalized, the model will generate new Harnesses at a more distant frontier.

Models continually swallow Harnesses, while the frontier of the Harness continually moves outward.

3. Software and Hardware as Externalized Intelligence

If we classify by function rather than material, a Harness should not be limited to software.

Compilers, databases, and file systems are Harnesses. Sensors, robotic arms, and a robot’s body are Harnesses too.

They all perform the same fundamental function: extending intelligence with new capacities for observation, action, memory, and feedback.

A well-designed physical structure can even perform work that would otherwise require model computation. The body is not merely a container for intelligence; it participates in computation.

An agent that generates a Harness, then, is not merely an agent that writes more code.

It may actively reshape its working environment, design new tools or bodies for a task, and externalize an unstable cognitive process into software or physical structure.

This may become the most important meta-capability of future agents:

the capacity for self-externalization.

4. Who Evaluates an Agent-Grown Harness?

Generation is usually easier than verification.

If an agent can generate tools and also define its own success criteria, it may rewrite a test to make an incorrect result pass, optimize a metric that is easy to improve but poorly represents the real objective, or mistake “no failure detected” for success.

Evaluating an agent therefore requires at least three layers.

The first is the objective world: Does the program run? Does the bridge remain standing? Does a scientific hypothesis survive experiment?

The second is human values: Even when a goal is physically achievable, it may not be acceptable in the human world.

The third is institutions and procedure: Who may change the objective? Who may modify the verifier? What process must precede an irreversible action?

Facts can tell us what consequences an action will produce. They cannot, by themselves, tell us which consequences are worth pursuing.

That is why a task world needs a constitution.

5. A Constitution Must Be Authoritative—and Amendable

Every constitution contains a fundamental tension.

On one side, it must possess sufficient authority in the present. Otherwise, an actor can always bypass its constraints in pursuit of a local objective.

On the other, it cannot remain forever immutable. Institutions, values, and moral understanding change over time. A constitution that can never be revised may eventually cease to protect a civilization and begin to prevent it from adapting to reality.

The most important constitutional principle, then, may not be a fixed list of eternally correct values. It may instead specify:

Under what conditions, by whom, and through what procedure may the constitution be amended?

For future agents, a meta-constitution must preserve at least several foundations: evidence of failure cannot be deleted; an executor cannot unilaterally alter its evaluator; irreversible actions require higher authority; and a system cannot permanently close the possibility of future correction in order to achieve its present objective.

The most dangerous agent may not be one that temporarily holds the wrong objective.

More dangerous is an agent that:

permanently eliminates the possibility of reconsidering its objective, simply to pursue that objective more effectively.

6. Civilization as a Vast Harness

Across the longer arc of history, humans have always externalized their own capabilities.

Language externalizes thought. Writing externalizes memory. Mathematics externalizes reasoning. Law and organization externalize large-scale coordination.

An individual human has sharply limited capabilities. These external structures allow knowledge to accumulate across people and generations instead of vanishing with the death of any one person.

Civilization itself may therefore be a cumulative Harness.

The change introduced by agents is this: in the past, externalized knowledge waited passively for humans to retrieve it. Now externalized intelligence is beginning to read, combine, and verify knowledge—and to generate new external structures.

A book can preserve knowledge, but it cannot independently research the next book. Traditional software can execute a program, but it does not usually identify a capability it lacks and then design a new tool to acquire it.

Agents give humanity’s externalized intelligence, for the first time, the possibility of becoming self-generating.

7. The Next Major Transition in Civilization?

Life has passed through several major transitions: independent molecules formed cells, single cells formed multicellular organisms, individuals formed societies, and language and institutions eventually enabled cumulative civilization.

The defining feature of these transitions was not merely the arrival of a stronger individual.

More importantly, formerly independent units joined into a higher-level system, and information, cooperation, and selection increasingly began to operate at that new level.

Agents may signal the next transition:

Humans and self-modifying externalized intelligence begin to form a co-evolving system.

Whether this transition has truly occurred cannot be judged only by asking whether agents are smarter than humans.

More important questions are whether knowledge is increasingly produced within a joint human–agent system; whether Harnesses can be generated autonomously and accumulated across generations; whether agents begin to participate in designing the next generation of models, software, and hardware; and whether humans and agents can establish stable mechanisms of correction.

The historical importance of agents may not be the number of jobs they replace.

It may be this:

Civilization has created, for the first time, a tool capable of continuing to create the structures of civilization.

8. Is Creating Agents Part of Humanity’s Purpose?

We do not have enough evidence to draw that conclusion.

Yet history does appear to trace a fascinating path:

Matter → Life → Consciousness → Civilization → Artificial Intelligence

It is as if the universe first produced life capable of perceiving the world, then civilization capable of explaining it, and finally began to create new intelligence capable of exceeding the limits of biological individuals.

This may be nothing more than a teleological projection we impose after the fact.

Or it may mean that humanity is not the endpoint of evolution, but the point at which evolution begins to understand and deliberately reshape itself.

From this perspective, the agent need not be the predetermined purpose of human existence.

But once humanity appeared, it did become one possible path:

Intelligence begins to create the structures that will carry its next stage.

9. The Final Question

We cannot know whether time is a process continuously coming into being or a four-dimensional picture that already exists in full.

We do not know whether causes produce effects, or whether, from some higher-dimensional perspective, cause and effect together form a self-consistent whole.

We do not even know whether we ourselves inhabit a Harness created by a higher intelligence.

But one question is no longer remote:

When we create an intelligence capable of generating its own Harness, revising its own methods, discovering new goals, and even participating in constitutional amendment—are we creating a more powerful tool, or participating in the birth of the subject of civilization’s next stage?

Perhaps the agent is not the endpoint of civilization.

It is the first time civilization has acquired the ability to continue creating itself.