Dual-System Graph Neural Cellular Automata for Autonomous Tool Dispatch and Schema Pruning

Abstract

In autonomous software agents, deciding which tool to invoke—such as reading a file, searching a database, or replying to the user—is fundamentally a micro-routing decision. Today’s systems dispatch every step through multi-billion-parameter Large Language Models (LLMs), spending seconds of latency and thousands of redundant tokens merely to select a function from a list. In this pa- per, I demonstrate that Neural Cellular Automata (NCAs) offer an ultra-lightweight alternative: a 59k-parameter network that functions as a fast reflex arc. Instead of generating tokens sequentially, my Graph Neural Cel- lular Automata (GNCA) framework parameterizes local cellular communication via continuous architecture meta-learning. Nodes representing user goals, execution history, and argument slots exchange local state updates over just 4 time steps to reach rapid consensus. I prove three structural theorems governing graph subsumption, receptive field diameter, and permutation symmetry, for- mally verifying all proofs in Lean 4 with zero unproven axioms. Extending training past 150,000 steps induces delayed generalization (grokking), lifting GNCA test top-1 accuracy to 68.07% (95% CI: [66.49,69.60]%), perform- ing within statistical parity of competitive tuned tabular ensembles (Random Forest 67.88%, HistGradientBoosting 69.47%) while running in sub-millisecond CPU latency (0.66 ms). Deployed as a System-1 micro-router paired with an LLM actuator in a Dual-System architecture, GNCA achieves an average routing latency of 21.43 ms across 89 live benchmark turns. This delivers a 78×–200× speedup over Open SOTA LLMs like Llama-4 Maverick and DeepSeek-V4. On the Sierra τ2-bench suite, the hy- brid agent reaches an 80.0% Pass1 reward with 100.0% write-action accuracy. On the Berkeley Function-Calling Leaderboard (BFCL v4), it delivers 90.0% tool irrelevance detection, eliminating schema distraction while cutting cumulative token consumption by 44.0% (92,272 tokens saved).

Keywords