Early agent harnesses were rigid scaffolds designed to force models along a specific path. As models become more intelligent and steerable, much of this scaffolding is no longer needed and can be deleted. The focus of modern harnesses is now on enabling longer, more complex execution chains.

The field is moving beyond labeling concepts with sparse autoencoders. The new frontier is understanding the intricate geometric structures (manifolds) these concepts form in a model's latent space and how circuits transform them, providing a more unified, dynamic view.

Research suggests a formal equivalence between modifying a model's internal activations (steering) and providing prompt examples (in-context learning). This framework could potentially create a formula to convert between the two techniques, even for complex behaviors like jailbreaks.

A pure 'pixels in, actions out' model is insufficient for full autonomy. Waymo augments its end-to-end learning with structured, intermediate representations (like objects and road concepts). This provides crucial knobs for scalable simulation, safety validation, and defining reward functions.

Moving beyond the simple Linear Representation Hypothesis, models organize concepts within sparse mixtures of subspaces. The specific geometry of these "manifolds" (e.g., a circle for days of the week) encodes the relationships and valid operations between concepts, like chemistry emerging from the periodic table.

Trying to simply block a model from learning an undesirable behavior is futile; gradient descent will find a way around the obstacle. Truly effective techniques must alter the loss landscape so the model naturally "wants" to learn the desired behavior.

Waabi's CEO explains that for physical AI, world models must go beyond just creating realistic simulations. The critical feature is 'controllability'—the ability to precisely generate and manipulate specific, safety-critical scenarios for testing. This is a fundamental difference from world models used for generating creative media or games.

Using a sparse autoencoder to identify active concepts, one can project a model's gradient update onto these concepts. This reveals what the model is learning (e.g., "pirate speak" vs. "arithmetic") and allows for selectively amplifying or suppressing specific learning directions.

When determining what data an RL model should consider, resist including every available feature. Instead, observe how experienced human decision-makers reason about the problem. Their simplified mental models reveal the core signals that truly drive outcomes, leading to more stable, faster-learning, and more interpretable AI systems.