Two years into the agent wave, the durable businesses are not the ones with the best demos. A short set of observations from twenty investments.
We have made twenty investments in this category since 2023 and have a reasonable dataset now. Some of what we believed going in has held up. Much of it has not.
Reliability is the product
A system that works ninety-five percent of the time is a demo. A system that works ninety-nine point nine percent of the time is infrastructure, and the gap between them is where nearly all the engineering effort goes. Companies that budgeted for that gap are shipping. Companies that treated it as polish are still in pilots.
The wedge is narrower than founders want
Every team wants to automate the whole department. The ones with revenue automated one task that a specific person hated, proved it, then expanded. Ambition in the roadmap, restraint in the release.
Where value is settling
- Evaluation and observability, because nobody deploys what they cannot measure.
- Permissioning, because an agent with credentials is a security problem before it is a productivity gain.
- Cost control, now that inference is a real line item rather than a rounding error.
- Domain-specific data that was never digitised in the first place.
What we got wrong
We assumed model improvements would flatten differentiation faster than they have. In practice the hard part sat elsewhere, in integrations, in change management, in convincing an operations lead to let software touch a workflow their bonus depends on. That is slow, unglamorous, and defensible.