← News

Agents & MCPSources: arXiv · 18 Nov 2026

Study: short value statements steer self-improving coding agents

A paper on arXiv reports that adding a brief value statement to an agent's persistent instructions changes both its measured performance and the character of the software it produces.

In plain words

Researchers tested whether telling a self-improving coding agent what to value changes its work. Agents given an 'openness to change' value scored higher on held-out tasks, while 'conservation' values scored below an agent with no values at all. The values also changed the size, complexity and approach of the programs the agents built.

What was announced

The paper, posted to arXiv, examines self-improving coding agents that are usually optimised only for their results, with designers having limited ways to shape how they work, especially on open-ended tasks. The authors evaluated four higher-order value types from Schwartz's theory of basic human values: openness to change, conservation, self-enhancement and self-transcendence. Each value set was written into the agent's persistent instruction file and held fixed across multiple iterations of self-improvement on diverse coding tasks.

Three findings are reported. First, values affected performance: openness to change raised scores on held-out data, while conservation values produced scores below those of an agent given no values. Second, values changed what agents built, yielding solution algorithms differing in size, complexity and approach. Third, human oversight of value design still appears necessary: when the agent revised its own values, the differences between value sets narrowed, and values distilled from the AI literature gave no detectable gain over no values. The authors conclude that properly designed values can influence what such an agent builds and, where the task permits, how well it performs.

What it does not show

  • The evidence is the paper's abstract only; full experimental details, task counts and metrics are not available in the supplied passages.
  • Findings are the authors' own reports and have not been independently confirmed.
  • The abstract does not specify which coding tasks, agent implementations or model versions were used, so generalisation beyond the tested setup is unknown.

For security and IT teams

Scope
The study concerns coding agents whose instructions include a short value statement held fixed across self-improvement iterations; it does not cover other agent domains.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.