agents IA

Ten times the human rate: that is the speed at which Joseph Valente and Michal Zalewski (Alphabet Security) say agents reach enterprise data, which is why their ACM Queue article moves the trust boundary from the application to the individual action. Eric Schmidt describes the same objects more plainly in the In Good Company podcast: a chain of query, reasoning, actions and synthesis, with almost any professional activity decomposable into concatenated agents.

Speed leaves a supervision problem behind. Rachel Laycock, CTO of Thoughtworks, accepts that AI produces more code than humans can reasonably review, and rejects the obligatory pull request as the remedy. Alexandre Frizzo, after a year of Claude Code, calls supervision the new bottleneck. Paula Hingel reads DORA 2025 along the same line: adoption correlates positively with throughput and negatively with delivery stability.

The counterweight proposed is scoping. Rakesh Gohel separates what is true from what matters now and what to do next, giving each its own graph. Ina Sebastian, Peter Weill, Thomas Haskamp and Jan vom Brocke sort decisions by ambiguity and risk: at One New Zealand, an agent translating marketing requests into SQL queries cut the time spent by 60%, with human review before launch, while any customer-service decision touching price keeps a human in the loop.

Aaron Levie goes furthest, treating agents IA as software's primary users: without an API, a CLI or an MCP server, a feature does not exist for them.