agents IA

Ethan Mollick's organizational-theory piece and Jamin Ball's systems-of-record analysis approach the same object from opposite ends: agents as coordination problem, agents as consumers of truth. That double framing recurs across the notes gathered here, and it is what makes "agents IA" less a product category than a stress test applied to existing structures.

The recurring move is measurement of what breaks. Joseph Valente and Michal Zalewski (Alphabet Security) build Beyond Zero on the observation that agents access data at ten times the human rate, which invalidates the three assumptions BeyondCorp rested on. Paula Hingel reads the DORA 2025 correlation the same way: adoption raises throughput and product performance while correlating negatively with delivery stability, a gap she attributes to uneven adoption without redrawn review boundaries. Alexandre Frizzo reports the same asymmetry from a single desk after a year with «Claude Code»: output up, supervision now the bottleneck, the writing muscle atrophied.

Aaron Levie pushes the consequence furthest, arguing that agents displace human users as software's primary consumers, which makes API-first infrastructure, CLIs, and «MCP» servers the condition of existing at all. Philippe Martin's BMAD assigns agents the roles of specification, architecture, artisans, and controllers, positioning method against «vibe coding» and the technical debt it accumulates.

The tension nobody here resolves: the same property that makes agents useful, acting across systems at machine speed, is the property that produces hallucinations, circular validation, and ambient authority.