What question did this study set out to answer?

The central aim is to reduce overhead in computer-use agents by implementing a routing system that minimizes reliance on large language models (LLMs).

May 14, 2026Open Access

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

Key Points

The central aim is to reduce overhead in computer-use agents by implementing a routing system that minimizes reliance on large language models (LLMs).
Developed kinthink, a four-layer router that performs TF-IDF matching and executes actions based on prompt hints.
Compared performance of routed agents using macbench tasks versus traditional LLM-agent methods.
Analyzed completion times and accuracy rates across 379 macOS-native tasks.
Routed agents achieved 48.0% task completion in 76 minutes, compared to 30.4% in 107 minutes for the LLM-agent baseline, indicating a 2x speed improvement.
Accuracy increased by 17.6 percentage points with the routed system, alongside a 99% reduction in token usage on the main task path.
In the web subcategory, 8 out of 10 tasks passed with an average time of 750 ms and zero token consumption.

Abstract

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.

Read Full Paperexternally

اسأل الذكاء الاصطناعي

Bookmark

View Full Paper