Today for AI
HOT RADAR
agentHEAT 19.3°

AI CURATED HOT STORY · 11/2/2026

Tsinghua & Microsoft Propose HarnessMM: Pre-planned Agent Steps Cut Token Use by ~90%

1 reports archived1 independent sourcesupdated 11/2/2026, 16:51:40
Synthesis & Latest Updates
1 Sources Cross-Validated

A joint study by Tsinghua University and Microsoft introduces HarnessMM, a framework where the model pre-writes future commands with validation checks. If tool results pass these checks, the agent executes directly without re-invoking the LLM. On HumanEval tasks, this approach reduced DeepSeek Harness's token usage by 87.93% while increasing success rates from 71.43% to 85.71%, significantly optimizing agent efficiency and cost.

LATEST/A joint study by Tsinghua University and Microsoft introduces HarnessMM, a framework where the model pre-writes future commands with validation checks. If tool results pass these checks, the agent executes directly without re-invoking the LLM. On HumanEval tasks, this approach reduced DeepSeek Harness's token usage by 87.93% while increasing success rates from 71.43% to 85.71%, significantly optimizing agent efficiency and cost.

Hourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 1 reports · Latest first
  1. Rohan Paul (@rohanpaul_ai)T2·78 pts

    Tsinghua & Microsoft Propose HarnessMM: Pre-planned Agent Steps Cut Token Use by ~90%

    • HarnessMM enables models to pre-plan command chains with success checks, skipping LLM decision-making if intermediate steps meet expectations.
    • On HumanEval benchmarks, applying this technique to DeepSeek Harness reduced token usage by 87.93% and increased success rate to 85.71%.
    • Pre-planning strategies significantly lower latency and compute costs for predictable task flows like coding, building, and testing.