# Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents

Source: [arXiv](https://arxiv.org/abs/2609.26760v1)  
Feed7 permalink: https://feed7.dev/p/2609-26760v1-0lklvan  
Published: 2026-09-22T17:40:45.000Z  
Trust: Needs Review (needs_review)

## Why Included

A learned harness moved recurring control into reusable code, cutting model calls and inference cost while keeping web-agent performance stable across 4B–120B deployment models.

## Source Summary

Growing Harness starts without a task-solving controller, then uses execution traces, joint repair, and rollback gates to accumulate reusable control code. Across two web-agent benchmarks and models from **4B to 120B**, it led **five of six** settings.

## Practical Implication

For recurring agent workflows, treat repeated planning and recovery logic as code that can be learned, tested, and retained. The reported harness reduced LLM calls by **76.0–91.8%** and inference cost by **74.4–98.6%** versus a tool-calling agent.

## Agent-Ready Context

Growing Harness starts without a task-solving controller, then uses execution traces, joint repair, and rollback gates to accumulate reusable control code. Across two web-agent benchmarks and models from **4B to 120B**, it led **five of six** settings.

For recurring agent workflows, treat repeated planning and recovery logic as code that can be learned, tested, and retained. The reported harness reduced LLM calls by **76.0–91.8%** and inference cost by **74.4–98.6%** versus a tool-calling agent.

The evidence comes from BrowseComp-Plus and WebArena-Verified, not general software work. The harness also requires trace-level localization, repair optimization, and held-out rollback checks; the abstract does not quantify their training cost.

## Connected Context

Feed7 judgment across 856 accumulated Signals:

This shifts harness engineering from manually specifying controls for each workflow to learning reusable control code from execution traces, with rollback checks limiting regressions. It gives benchmark evidence that retained scaffolding can reduce calls and cost, while narrowing the claim to two web-agent settings and leaving harness-training cost unquantified.

- [TurnSight: Turn-Level Hindsight Self-Distillation for Tool-Integrated Reasoning](https://feed7.dev/p/2608-04007v1-0gjxuy2) — TurnSight supplies turn-level credit assignment for execution traces, while Growing Harness uses traces to localize and repair reusable control code.
- [Twin: Playing an Unknown Game with a Test-Time Digital Twin](https://feed7.dev/p/2608-14490v1-0d3xjvt) — Both retain executable machinery learned during interaction and gate changes through validation; Twin validates simulated transitions, whereas Growing Harness applies held-out rollback checks to control-code repairs.
- [Stop AI Agent Hallucinations: 5 Techniques + Production Patterns - Elizabeth Fuentes, AWS](https://feed7.dev/p/stop-ai-agent-hallucinations-5-techniques-production-patterns-elizabeth-09g1w9s) — The AWS patterns describe manually encoded runtime controls; Growing Harness suggests repeated planning and recovery controls can instead be accumulated and tested as reusable code.

## Context Map

- Layer: agent
- Domains: research
- Topics: harness-engineering, agent-reliability, tool-use

## Uncertainty

- The evidence comes from BrowseComp-Plus and WebArena-Verified, not general software work. The harness also requires trace-level localization, repair optimization, and held-out rollback checks; the abstract does not quantify their training cost.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.