AI agents just hit a wall, and it isn't capability

The newest measurements give agents a real ceiling. They hit 50% success on well-specified, multi-hour tasks, roughly what a new hire could do without any institutional knowledge. The moment the work needs organisational context and several stakeholders, performance drops sharply.

Growth and product loops are precisely that kind of work: cross-functional, multi-stakeholder, context-rich. So the advantage doesn’t go to whoever automates their loops first. It goes to whoever maps the line between agent and human most precisely.

That reframes the whole investment. The bottleneck isn’t agent capability, which keeps improving. It’s your organisation’s capacity to keep redrawing what agents can and can’t handle as the ceiling rises. Treat agent automation as a scoping problem, not a capability race.