aifollow.news
← TrendingLatest development Nov 20 05:05
STORY IN FOCUS

Google research: Agent framework edits raise coding task success from 5% to 40%

Google researchers proposed letting an optimization model edit one module of an AI agent's operating framework at a time, retaining changes only after validation. In a SWE-bench Lite test using Gemini 3.8 Flash and 40 tasks excluded from optimization, success rose from 5.0% to 40.0%, with about 46,700 tokens used per task on average. The results are limited to specific models and software engineering tasks and do not establish the same gains across all agent applications.

1 developments2 reports2 sources

Event developments

Coverage of the same occurrence shares a node; subsequent developments have their own nodes.

Google paper tests agent harness improvements through gated module edits

An optimizer model edited one of four modules in a 142-line coding agent harness at a time, with each change screened by a validation set and token cap. On 40 held-out SWE-bench Lite tasks with Gemini 3.8 Flash, the gated harness solved 40.0% using 46.7k tokens per task; an unchecked full rewrite solved 42.5% using 451.6k tokens per task.

View source reports · 2 sources · 2 reports
All reports · 2

By original publication time, with each report's bookmarks and feedback preserved.

IT之家 科技新闻 Machine translation

Google research: Agent framework edits raise coding task success from 5% to 40%

Google researchers proposed letting an optimization model edit one module of an AI agent's operating framework at a time, retaining changes only after validation. In a SWE-bench Lite test using Gemini 3.8 Flash and 40 tasks excluded from optimization, success rose from 5.0% to 40.0%, with about 46,700 tokens used per task on average. The results are limited to specific models and software engineering tasks and do not establish the same gains across all agent applications.

Rohan Paul (@rohanpaul_ai) Machine translation

Google paper tests agent harness improvements through gated module edits

An optimizer model edited one of four modules in a 142-line coding agent harness at a time, with each change screened by a validation set and token cap. On 40 held-out SWE-bench Lite tasks with Gemini 3.8 Flash, the gated harness solved 40.0% using 46.7k tokens per task; an unchecked full rewrite solved 42.5% using 451.6k tokens per task.