Source note
The Usefulness of AI Agents
Summary
The essay argues that AI agents have clear value in coding work, but their broader usefulness is narrower than current adoption pressure suggests. It frames agents as speed tools that can reduce some digital labor while increasing oversight, ethical, and human-agency costs.
Problem
- The piece addresses rushed AI-agent adoption by managers, researchers, and developers without a clear account of value, cost, or human control.
- It matters because agents can speed up software and research workflows while also producing low-value output, increasing token use, and reducing human involvement in decisions.
- The author also raises risks around copyright, labor exploitation, energy use, and the loss of open online sharing when AI systems can reuse public material at scale.
Approach
- The author uses a reflective essay format rather than a controlled experiment or benchmark study.
- He compares personal use across everyday life, writing, coding, and research tasks after about 2 years of work on LLM-powered agents.
- For software work, he reports testing GitHub Copilot, OpenAI Codex, Claude Code, and Goose with different LLM backends.
- He proposes restrictive coding-agent instructions: change at most 1-2 files per request, keep edits focused, avoid unrequested refactors, state a plan before coding, and seek approval for changes over 100 lines or multiple files.
- For research automation, he describes using OpenAI Codex with GPT-5.4 and extra-high reasoning effort to collect data, analyze it, draft discussion, and compile a paper-like PDF.
Results
- The excerpt reports no benchmark metrics, controlled comparisons, or quantitative breakthrough results.
- In fall 2025, the author found coding agents unwieldy: they produced too much code and made projects harder to manage; he preferred inline autocomplete at that time.
- By the time of writing, he says coding agents had improved enough to build smaller prototypes and projects while preserving his desired oversight.
- With the restrictive prompt rules, coding agents became more usable for his workflow, especially when limited to 1-2 files and gated before changes above 100 lines.
- In research tasks, Codex with GPT-5.4 produced coherent papers, analysis scripts, figures, tables, discussion, and PDFs without human intervention, but the author judged the output as uninteresting or weak as research.
- The strongest claim is qualitative: AI agents can make digital work faster, especially programming and data-analysis setup, but the essay finds limited evidence that they create better judgment, better research value, or better quality of life.