10 ideas from AI Engineer on how development changes when agents write code
OpenAI
Anthropic
GitHub
OpenAI conducted an experiment where three engineers created a product of about a million lines of code in five months, with all code generated by Codex, in about a tenth of the traditional time. The key bottleneck shifted from code production to the ability to specify what to build, organize autonomous agents, verify results, and prevent system destruction. These insights were presented at AI Engineer 2026 conferences in London and San Francisco, with talks on harness engineering, long-running agents, multi-agent coordination, and more.
At the AI Engineer 2026 conferences in Europe and World's Fair, several talks explored how software development changes when AI agents write code. OpenAI's experiment with Codex, where three engineers built a million-line product in five months without writing a single line manually, showed that code generation speed is no longer the bottleneck; the new constraint is human ability to specify requirements, organize autonomous agents, verify output, and prevent accelerated system collapse. Ryan Lopopolo introduced 'harness engineering': a framework where humans steer and agents execute, with infrastructure including documentation, tests, architecture rules, and feedback mechanisms. Ash Prabaker and Andrew Wilson from Anthropic addressed long-running agents, proposing external state management, adversarial evaluation, and handoff procedures to maintain coherence over hours. Maggie Appleton from GitHub Next discussed the social dimension of multi-agent systems, warning that two dozen agents do not automatically form a team and that coordination, conflict resolution, and focused design are needed to avoid chaos.
- Сокращения
- CI = Continuous Integration — непрерывная интеграция
- DOM = Document Object Model — объектная модель документа
- JSON = JavaScript Object Notation — формат обмена данными
Source: Habr — хаб ИИ —
original