Skip to content
Trending storyDeveloping

HarnessMM cuts agent token use by reusing pre-planned command chains

2 articles2 sourcessince Oct 26Last article 10h ago ·

Overview

AISummary of 2 articles

Researchers from Tsinghua and Microsoft Research Asia propose HarnessMM, a method that lets an agent run pre-planned command chains without calling the LLM each step.

The model writes future commands with exact arguments and a check for each, such as a successful build. When a tool result passes the check, the agent runs the next command and skips the LLM call. The stored decisions sit in an Executable Decision Graph with actions, arguments, dependencies and guards.

On 98 HumanEval tasks, Elvis Saravia (@omarsar0) reports that a DeepSeek harness cut tokens per task from 96.28k to 11.62k while success rose from 71.43% to 85.71%. Rohan Paul (@rohanpaul_ai) reports the same DeepSeek results as an 87.93% token reduction. With Codex, success stayed at 90.82% while tokens per task fell 29.54%. A separate recursive self-improvement loop tunes the harness configuration from training feedback without changing model weights, and the Codex result is attributed to that loop.

Written by AI from the articles below · updated Nov 2, 4:28 AM ET

Check the sources:

Latest developmentNov 2, 3:51 AM ET
HarnessMM cuts agent token use by reusing pre-planned command chains

Developments

2 developments

  1. Nov 2, 3:51 AM ET · 1 article
    Tsinghua and Microsoft paper introduces HarnessMM, cutting tokens for coding agents
    Rohan Paul: HarnessMM cuts agent token use by reusing pre-planned command chains
  2. Oct 26, 10:53 AM ET · 1 article
    HarnessMM cuts agent harness tokens per task via cached decisions and RSI-tuned configuration
    Elvis Saravia: HarnessMM cuts agent tokens per task by 88% with cached decisions

Article timeline

The articles in this story. Times are ET.

Nov 2
  1. Rohan PaulX
    HarnessMM cuts agent token use by reusing pre-planned command chains

    AIA Tsinghua and Microsoft Research Asia paper introduces HarnessMM, which lets a model plan future agent commands with pass checks so the agent can skip LLM calls when a check succeeds. On 98 HumanEval tasks, Codex kept 90.82% success while cutting tokens per task 29.54%, and DeepSeek Harness rose from 71.43% to 85.71% success while cutting tokens 87.93%.

    Image from @rohanpaul_ai's post
Oct 26
  1. elvisX
    HarnessMM cuts agent tokens per task by 88% with cached decisions

    AIResearchers propose HarnessMM, which stores settled agent decisions in an Executable Decision Graph and reuses them when guards hold. On a DeepSeek harness, tokens per task fall from 96.28k to 11.62k while success rises from 71.43% to 85.71%. A recursive self-improvement loop tunes the harness configuration from training feedback without changing model weights, and with Codex on HumanEval, success stays at 90.82% while tokens per task drop 29.54%.

    Image from @omarsar0's post

Heat Trend

Current comparable heat
8
Peak comparable heat
10Nov 2, 5:00 AM ET
Change in past 24 hours
–

The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.