Google study: self-modifying agents improve task success rate 8x with harness
11/20/2026 — 11/20, 09:19·1 sources (all)·1 reports (all)
Key takeaways
- Google researchers published a paper titled 'Harnesses That Improve Themselves' on November 17, letting AI agents iteratively modify their o…
- the study also found that a harness optimized for one model loses effectiveness when transferred to another.
Background & analysis
According to a report published by IT Home on November 20, 2026, researchers at Google released a paper on November 17, 2026 titled Harnesses That Improve Themselves. The work describes a setup in which an AI agent incrementally modifies the harness that wraps it, then verifies each modification before keeping it, with the goal of raising performance on software engineering tasks.
The study centers on the AI agent harness rather than the underlying model itself. A harness is the operational framework built around a model, responsible for managing context and memory, controlling multi-turn execution, invoking external tools, and validating results before a task is closed. In tests on SWE-bench Lite, an agent using the modular optimization approach raised task success from 5.0% to 40.0%, an 8x improvement, at an average cost of roughly 46,700 tokens per task.
The researchers also note that models can differ noticeably in how they use tools and where their weaknesses appear during execution. A harness tuned for one model may therefore stall, consume excessive tokens, or perform worse when transferred to another model. The paper proposes a self-improving harness architecture as a response, moving beyond the common practice of constraining agents purely through prompt edits. The available coverage does not report follow-up validation of the architecture under cross-model transfer, so the account currently stops at the paper's release and the benchmark results.
AI-generated from 1 reports · updated 2 days ago
- Heat index
- 25
- All sources
- 1
- All reports
- 1
- First seen
- 2 days ago
Reports on this story headlines open the original
Google researchers published "Harnesses That Improve Themselves" on November 17, letting AI agents iteratively modify their own runtime harness and verify each change. On SWE-bench Lite, the modular approach raised task success from 5.0% to 40.0% at about 46.7k tokens per task; the paper also notes harnesses tuned for one model can stall or degrade when moved to another.
IT之家AI score 60
Other stories people are talking about
- Heat index 1143RisingAnthropic Reaches $1.6 Trillion Market Value in First Trading Week After IPO21 sources
- Heat index 1011Anthropic Mathematician Says Claude Formalized the Generalized Riemann Hypothesis21 sources
- Heat index 622Anthropic IPO total raise reaches $93.5 billion after greenshoe exercise20 sources
- Heat index 596RisingThinking Machines Lab Open-Sources Inkling 2 Mini with 32GB Model Card Recommendation10 sources
- Heat index 490RisingSEC probes Nvidia accounting practices as shares post seven-week slide9 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion

