Agents de codage

Since March 2026, hours worked by agents at Anthropic have exceeded hours worked by humans, and commits per engineer there multiplied by five in six months. Agents de codage, the AI agents that write and modify code, are what that count measures. Claude Code and Codex belong to the category, which itself sits inside the broader class of computer-use agents. Martin Monperrus cites SWE-bench rates of roughly 20 to 40 percent of real issues resolved; David Heinemeier Hansson says frontier models, especially when combined, now surpass virtually every programmer across a broad spectrum of skills.

The bottleneck moved, and the sources disagree about where it landed. Monperrus argues in a June 2026 position paper that human code review is no longer a necessary stage, and proposes multi-agent adversarial verification in its place. Addy Osmani puts the constraint at verification itself: an agent that cannot check its own work guesses. Dorra Bartaguiz took a single-prompt generation announced as 100 percent covered and measured 33 percent global coverage with a mutation score of 13.82 percent. Patrick Debois locates it upstream, in context treated as a versioned, tested artifact.

The recorded costs are specific. Roughly 80 percent of agent session time at Anthropic is spent waiting for a human reply, and meta-work rose from 10 to 22 percent of sessions. Osmani names the joy of knowing as the most threatened: choosing among an agent's proposals is a different skill from designing one, and the capacity to design erodes when only choosing gets practised.