Grep-Routed Agents: Bypassing the LLM Tax on Computer-Use Tasks

Computer-use agents conventionally consult an LLM on every action: read the prompt, decide which tool, format arguments, execute, read result, decide next step. We observe that for ~80% of macbench's 379 macOS-native tasks, those LLM round-trips are pure overhead — the natural-language prompt already implies one canonical shell action (a cerebellum action), and the work of choosing it can be done by grep against a small index. We present kinthink, a four-layer router that extracts Fast-path hints, does TF-IDF matching against 239 prompt examples, substitutes slot values, and executes the matched cerebellum action — all in 6-25 ms of shell, consuming zero LLM tokens on the hit path. On macbench 379 tasks: 48.0% pass in 76 min vs 30.4% in 107 min for the unrouted LLM-agent baseline (2x speed, +17.6pp accuracy, 99% token reduction on the dominant path). Web subcategory: 8/10 pass at 750ms avg / 0 tokens — direct counter to OpenAI Codex Chrome Extension (2026-05-07). Fourth installment in the LocalKin thesis chain: paper #1 (retrieval), #5 (cognition), #10 (LLM-tax measurement), this paper (routing) — all showing that for bounded domains, deterministic shell beats LLM round-trips.