The Agent Computer

The Agent Computer is the runtime Jax executes on — the operating system for a spatial agent. It pairs a fast, deterministic compute layer with verifiable optimization solvers, so the agent can reason about the physical world and prove its answers.

Core concepts

A language model alone cannot plan a network or route a fleet. It reasons in tokens; the physical world runs on geometry and constraints. The Agent Computer is the layer that bridges the two: it gives Jax dedicated, deterministic compute that a model can call but cannot fudge.

• Deterministic compute. Geo-math — distance, centroid, snap-to-roads, route-along-roads, isochrone — is computed exactly, the same way every time, through the Run primitive. The answer does not depend on the model's mood.
• Verifiable optimization. Hard decisions — siting, design, routing, scheduling — are dispatched through the Solve primitive to a family of solvers that return plans which are provably optimal, with an optimality gap to prove it.
• One world model underneath. Both compute layers read their inputs from, and write their results to, the DataStore. The agent never passes large blobs of data inline; it points compute at files.

In marketing terms this layer is the Agent Computer. In the platform, the optimization engine is the MathEngine and deterministic geometry runs on the GeoEngine. They are the same runtime, named for what they do.

How it works

The agent runs a read-reason-write loop. Every step is one of the eight primitives, and the heavy lifting happens in the compute layer.

1. Read. Jax reads the relevant slice of the world model — a polygon of demand, a set of assets, the active constraints.
2. Reason. The model decides what to compute: a route, a layout, a schedule, a simulation.
3. Compute. Run returns exact geometry in milliseconds; Solve formulates the problem, dispatches it to the right solver, and returns a verified plan.
4. Write. The result is written back to the DataStore as a normal versioned change — reviewable, mergeable, reversible.

Verifiable, not plausible

The point of the Agent Computer is trust. A plausible answer from a language model is not good enough to commit to live infrastructure. Every decision the Agent Computer makes carries evidence:

• Optimality gap. Solve reports how close a result is to provably optimal, so you know exactly how much trust to place in it.
• Hybrid dispatch. Small problems solve instantly in the browser; large ones dispatch to a dedicated solver server. The engine routes each problem to the right place automatically.
• Self-correction. When a model is infeasible or weak, the engine diagnoses why and reformulates rather than returning a silent failure.
• Auditable record. The full formulation — objective, variables, constraints — and the result are retained alongside the plan, giving a defensible trail.

Why a dedicated runtime

Bolting optimization onto a chatbot produces a demo, not a system. The Agent Computer is a real runtime because spatial decisions demand it:

• Determinism so the same inputs always produce the same geometry.
• Reproducibility so a plan can be re-derived and audited later.
• Scale so a national program's problems still finish, not just a single site's.

This is what makes spatial intelligence feel instant and stay trustworthy — the model proposes, the Agent Computer proves.

Where to go next

• Understand the tool surface: The eight primitives.
• Go deep on optimization: MathEngine.
• Go deep on deterministic geometry: GeoEngine.
• See the agent end-to-end: Jax overview.