All insights

Quarterly review18 June 20251 min read

Coding agents go mainstream

Claude 4, OpenAI's o3 and Codex, and Google's I/O announcements made AI agents that carry out whole tasks part of everyday software work.

What happened this quarter

April brought Meta's Llama 4, OpenAI's GPT-4.1, and OpenAI's o3 and o4-mini reasoning models. In May, OpenAI previewed Codex, an agent that works on software tasks in the cloud, and Anthropic released Claude Opus 4 and Claude Sonnet 4 alongside general availability of Claude Code. At Google I/O, Google showed Veo 3, which generates video with sound, and expanded its Gemini 2.5 models.

The Model Context Protocol, published by Anthropic last November, also gained wide support: OpenAI, Google and Microsoft all announced that their tools would work with it.

Our take

For the first time, mainstream tools let an AI agent take a task such as "add this field, update the tests, fix what breaks" and work on it for many minutes without a person typing each step. This is the way of working we have been building towards, and we now use it every day.

It also changes where the effort goes. Less time goes into typing code; more goes into describing the task well, reviewing the result and checking that tests cover what matters. A senior person who can do those things well gets far more done. A process without them produces more code, faster, with no one sure whether it is right.

Wide support for MCP is good news for buyers. A connection built once can now work across tools from several vendors.

Keep reading

Related insights