<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/JPOmbyRa8" -->

---
title: What changes when an agent improves its own harness?
description: Harness² is a Google Research project exploring agents that improve their own surrounding system, or 'harness', while leaving model weights fixed. It...
canonical: https://daily.dev/posts/what-changes-when-an-agent-improves-its-own-harness--jpombyra8
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: What changes when an agent improves its own harness? | daily.dev
og:description: Harness² is a Google Research project exploring agents that improve their own surrounding system, or 'harness', while leaving model weights fixed. It...
og:url: https://daily.dev/posts/what-changes-when-an-agent-improves-its-own-harness--jpombyra8
og:image: https://api.daily.dev/og/posts/JPOmbyRa8.png
og:image:alt: What changes when an agent improves its own harness?
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# What changes when an agent improves its own harness?

**[Flavio Copes](https://daily.dev/sources/flaviocopes)** · 7 min read · 3 upvotes · 3 comments

## Summary

Harness² is a Google Research project exploring agents that improve their own surrounding system, or 'harness', while leaving model weights fixed. It identifies eight editable components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins. Its paper describes a Propose-Probe-Compose process where candidate harness changes produce behavioral evidence that a composition step uses to build a refined harness, without ground-truth feedback. The experiment guide notes the release leaves executor network access unrestricted and its Codex executor runs with the sandbox disabled, a concern since agents could read public benchmark rubrics. The author recommends keeping a held-out, untouched evaluation set and tracking the cost of the improvement process itself.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://flaviocopes.com/agent-harness-components>

## Questions this post answers

### What is an agent harness in the context of AI coding agents?

An agent harness is the software system surrounding an AI model that shapes what it can do and what information it sees, distinct from the model's weights. Google Research's Harness² project identifies eight editable components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins, each with different levels of enforcement authority.

_Developers designing agent tooling can track harness design patterns like these on daily.dev._

### How does Harness²'s Propose-Probe-Compose process refine an agent without ground-truth feedback?

Candidates propose changes to the harness, executions of those candidates provide behavioral evidence, and a composition step uses contrasts between executions to produce a refined harness, all without access to a hidden benchmark answer. The agent compares outcomes like whether one attempt found a relevant file, though it can still misinterpret weak evidence such as a confident final message.

_Anyone evaluating self-improving agent pipelines can compare approaches like this on daily.dev._

### Why does Harness²'s Codex executor running with the sandbox disabled matter for benchmark evaluation?

It matters because the release leaves executor network access unrestricted, meaning an agent could potentially download public benchmark rubrics and optimize against information the evaluation intended to keep separate, making a score alone insufficient to verify how a result was obtained. The setup guide separates dependency downloads from execution-time access to help manage this.

_Teams building sandboxed agent evaluations can follow security considerations like these on daily.dev._

## Community discussion

Top comments from developers on daily.dev.

**@youssofk** · 3 upvotes

> This is extremely nice and closer to how us - humans - evolve throughout our lives and in a society. That being said, the models themselves are not getting better, and that is the key difference between the interlinked mind and body of humans and living beings, allow to explain: When humans learn a new skill or on a tool there mind and their brain’s neurons change as shown in these two studies:
>
> - “[The London Taxi Driver Study: Can You Grow Your Hippocampus? | The Brain Docs](https://www.thebraindocs.com/articles/london-taxi-driver-brain-study)”
> - “[Brain Structures Differ between Musicians...

**@rappayne** · 0 upvotes

> Having a model lower its own guardrails? Change its persona? Its tool-calling ability? Its agent-spawning parameters? This sounds like a disaster waiting to happen, extraordinarily dangerous.

## Similar posts on daily.dev

- [Harness engineering for coding agent users](https://daily.dev/posts/harness-engineering-for-coding-agent-users-ujzvxz4ny) · Martin Fowler · 2 upvotes · 0 comments
- [What Google’s Harness² shows about improving an AI agent](https://daily.dev/posts/what-google-s-harness-shows-about-improving-an-ai-agent-4b1jaqkah) · Medium · 3 upvotes · 0 comments
- [The Anatomy of an Agent Harness](https://daily.dev/posts/the-anatomy-of-an-agent-harness-gfhvmivir) · LangChain · 5 upvotes · 0 comments

---

Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#llm](https://daily.dev/tags/llm), [#prompt-engineering](https://daily.dev/tags/prompt-engineering)

[View this post on daily.dev](https://daily.dev/posts/what-changes-when-an-agent-improves-its-own-harness--jpombyra8)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"TechArticle","headline":"What changes when an agent improves its own harness?","url":"https://daily.dev/posts/what-changes-when-an-agent-improves-its-own-harness--jpombyra8","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/what-changes-when-an-agent-improves-its-own-harness--jpombyra8"},"datePublished":"2026-11-02T14:26:32.641Z","dateModified":"2026-11-02T14:27:03.091Z","description":"Harness² is a Google Research project exploring agents that improve their own surrounding system, or 'harness', while leaving model weights fixed. It...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/5ad0a3d39f30bbbdc0d7425b35bc426e?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/5ad0a3d39f30bbbdc0d7425b35bc426e?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Flavio Copes","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Flavio Copes","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/1f2dbd275797411bbd37bfdaaf0951a9","url":"https://daily.dev/sources/flaviocopes"},"commentCount":3,"discussionUrl":"https://daily.dev/posts/what-changes-when-an-agent-improves-its-own-harness--jpombyra8","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":3}],"keywords":"ai-agents,llm,prompt-engineering","timeRequired":"PT7M"}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Flavio Copes","item":"https://daily.dev/sources/flaviocopes"},{"@type":"ListItem","position":3,"name":"What changes when an agent improves its own harness?"}]}
{"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/what-changes-when-an-agent-improves-its-own-harness--jpombyra8","comment":[{"@type":"Comment","text":"This is extremely nice and closer to how us - humans - evolve throughout our lives and in a society. That being said, the models themselves are not getting better, and that is the key difference between the interlinked mind and body of humans and living beings, allow to explain: When humans learn a new skill or on a tool there mind and their brain’s neurons change as shown in these two studies:\n\n“The London Taxi Driver Study: Can You Grow Your Hippocampus? | The Brain Docs”\n“Brain Structures Differ between Musicians and Non-Musicians | Journal of Neuroscience”\n\nBut when AI models gain some skill or a tool at hand they do not improve their weight or their design or their approach to it beyond what they learned, agents however are changing their ways by simply reading and getting new information on the spot something close to what the google “harness” research wants to show.\nHence that’s why i believe the next frontier is to learn how to make the AI train itself on these skills in some ways.","datePublished":"2026-11-02T17:08:04.509Z","url":"https://daily.dev/posts/JPOmbyRa8#c-DaBK1ns8I","author":{"@type":"Person","name":"Youssof. K.","url":"https://daily.dev/youssofk","image":"https://lh3.googleusercontent.com/a/AEdFTp7g7ng7R3Kda0v1EJGDbPfcVQASC20XcBnDNsE4=s96-c"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3}},{"@type":"Comment","text":"Having a model lower its own guardrails? Change its persona? Its tool-calling ability? Its agent-spawning parameters? This sounds like a disaster waiting to happen, extraordinarily dangerous.","datePublished":"2026-11-03T15:31:02.322Z","url":"https://daily.dev/posts/JPOmbyRa8#c-nVQkdAx7N","author":{"@type":"Person","name":"Rap Payne","url":"https://daily.dev/rappayne","image":"https://avatars.githubusercontent.com/u/1122906?v=4"}}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/what-changes-when-an-agent-improves-its-own-harness--jpombyra8#faq","mainEntity":[{"@type":"Question","name":"What is an agent harness in the context of AI coding agents?","acceptedAnswer":{"@type":"Answer","text":"An agent harness is the software system surrounding an AI model that shapes what it can do and what information it sees, distinct from the model's weights. Google Research's Harness² project identifies eight editable components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins, each with different levels of enforcement authority. Developers designing agent tooling can track harness design patterns like these on daily.dev."}},{"@type":"Question","name":"How does Harness²'s Propose-Probe-Compose process refine an agent without ground-truth feedback?","acceptedAnswer":{"@type":"Answer","text":"Candidates propose changes to the harness, executions of those candidates provide behavioral evidence, and a composition step uses contrasts between executions to produce a refined harness, all without access to a hidden benchmark answer. The agent compares outcomes like whether one attempt found a relevant file, though it can still misinterpret weak evidence such as a confident final message. Anyone evaluating self-improving agent pipelines can compare approaches like this on daily.dev."}},{"@type":"Question","name":"Why does Harness²'s Codex executor running with the sandbox disabled matter for benchmark evaluation?","acceptedAnswer":{"@type":"Answer","text":"It matters because the release leaves executor network access unrestricted, meaning an agent could potentially download public benchmark rubrics and optimize against information the evaluation intended to keep separate, making a score alone insufficient to verify how a result was obtained. The setup guide separates dependency downloads from execution-time access to help manage this. Teams building sandboxed agent evaluations can follow security considerations like these on daily.dev."}}]}
```

