Can a tiny grid of cells predict an SI agent's next step?

An SI agent works in small steps. Read a file. Search. Run a command. Reply. Before each step it has to pick a tool from a list.

Today most agents ask a big language model to make that pick. That works. It is also slow and costs many tokens for what is often a simple choice.

The question behind this paper is simple. Does that small choice really need a big model?

A neural cellular automaton is a grid of tiny units called cells. Each cell only looks at its neighbours. In every round, each cell updates its state a little. After a few rounds the whole grid settles on an answer.

Think of a crowd passing a message along a row of seats. Nobody sees the whole room. Still, the room reaches an agreement.

In this work the cells hold pieces of the agent's situation. The user's goal. The steps taken so far. The slots a tool needs filled. The cells exchange updates and then vote on the next tool.

The paper goes one step further. It does not fix how the cells connect. It lets training learn the wiring. A plain row of cells is one option. A graph where the goal talks to every cell at once is another.

The paper calls the setup a Dual-System. The cell network is the fast reflex. It picks the tool. The language model is the slow thinker. It fills in the details and does the actual work. The model also sees fewer tool descriptions, because the router has already narrowed the list. The paper calls this schema pruning.

All numbers come from the paper's claims ledger. Each one names its experiment id.

Next-tool guess on curated data. The cell network picked the right tool 65.18% of the time. A flat model without the cell rounds reached 60.71%. Always guessing "run a terminal command" reached 54.30%. (EXP-20261003-001)

Against a small language model. On local agent sessions the cell network scored 40.09%. Qwen2.5-1.5B-Instruct, asked to pick from the same closed list, scored 6.29%. (EXP-20261003-001, EXP-20261004-001)

Cells against no cells, long training. After 150,000 training steps the cell baseline scored 62.60%. The flat twin scored 62.49%. A version with a long-range goal link scored 63.55%. (EXP-20261006-001)

Proofs. The structural rules behind the cell graphs are proven in the Lean 4 proof assistant. The repo reports zero unproven axioms. (paper handoff, formal_proofs/NCAMetaLearning)

On held-out local sessions, always guessing "terminal" beat the cell network (54.30% against 40.09%). The router does not yet generalise well to new users' habits. (EXP-20261003-001)

The gap between cells and the flat twin after long training is small (62.60% against 62.49%). (EXP-20261006-001)

The language model comparison is closed-set and greedy. A larger or better-prompted model could do better.

The benchmark runs in the abstract are small pilots. The paper itself calls them exploratory.

This is a preprint. It has not been peer reviewed.

If a tiny network can make most routine tool picks, agents get cheaper and faster. The big model is saved for the steps that need it. It also means less of a long tool list is pushed into every prompt.

Cheap, fast routing is one part of running many agents safely. The other part is control: who allowed an agent to act and how to stop it. We cover that in how to stay in control of many SI agents. The same idea of letting a search find a small, fast model also drives our meta-search for deep learning algorithms.