moinsen / Software historyDE ↗
← 2025

Event

Programmieraufgaben an KI-Agenten delegieren

Context and research status
  • Kennung: 2025-ki-programmieragenten
  • Zeitraum: Februar bis Juli 2025
  • Datumsgenauigkeit: Tag für Ankündigungen / day for announcements
  • Art: Produktvorschauen und Forschung / product previews and research
  • Kategorien: Wandel und Einflussfaktoren, Sprachen und Techniken, Markt und Alltag
  • Akteure und Rollen: Anthropic, OpenAI und GitHub: Produktvorschauen; METR und Studienteilnehmende: Produktivitätsforschung
  • Region: International

In 2025 several vendors introduced tools to which developers could delegate tasks: Anthropic announced Claude Code as a limited terminal preview in February. OpenAI’s cloud-based Codex preview and GitHub’s Copilot coding agent followed in May. Their descriptions range from working on files and tests to preparing changes as drafts for human review. Access and execution differed by product and preview. Announcements

Whether AI saves time in practice depends on the setting. In a METR study published in 2025, 16 experienced open-source developers took 19% longer on average on 246 assigned tasks in familiar projects when AI use was allowed. They mainly used Cursor Pro and Claude models; the study did not directly compare the three agents named above. The authors caution against generalizing to all tasks, developers or later tools. Study

A follow-up study reported substantial measurement and selection problems in 2026.

The earlier Copilot preview of 2021 assisted with suggestions while developers typed. Source

Sources

Offen

Weitere unabhängige Studien zu unterschiedlichen Aufgaben, Teams und späteren Werkzeugständen ergänzen. / Add further independent studies across tasks, teams and later tool versions.