Coding agents improve software speed by revising their own setup

Tan and colleagues report that agents allowed to revise their operating software during a task achieved higher correctness-preserving speedups than agents using fixed or externally revised setups.

Big Tech

Yuhao Tan · Hanyun Jiang · Fangkai Yang · Yuan Yao · Taolue Chen · Bo Qiao · +5 more

Nanjing University · Zhejiang University · Birkbeck, University of London · Microsoft

Research Digest·9 November 2026·2 min read

Illustration: Coding agents improve software speed by revising their own setup

In one line

Agents that improve their own code permission harnesses during a task achieve higher performance speedups than offline optimization.

Before you read this

Three ideas this paper builds on

agent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.173 papers →recursive self-improvementAn AI system repeatedly uses its own outputs to enhance its capabilities, creating a feedback loop for autonomous improvement.58 papers →test-time adaptationAdapting a deployed model during inference using new data or interactions, without a separate training phase.81 papers →

Read it as

When a software agent runs a long task, the code that controls what it can do and see, its harness, might need fixing. Previously, harnesses were improved only between tasks, using past data. This paper studies whether an agent can fix its own harness while the task is still running.

The authors built a system called BoxBox that lets the agent pause its work, review what went wrong, and rewrite the harness during the task. The new harness is loaded instantly while the agent keeps its place. They also created a benchmark called PerfRace for testing this on real software optimization jobs.

On PerfRace, BoxBox achieved a speedup of 1.70 times after four hours, outperforming the best existing approach at 1.54 times. This shows that letting agents improve their own tools mid-task can make them more efficient and cost-effective.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

agent harness recursive self-improvement test-time adaptation self-evolving agent coding agent harness evolution