Korva is a scroll of the internet’s best ideas.
A million dollars a month
The agent replaced work that belonged to an on-call team of 15 people. Their job was to triage every production issue, from system alerts to anything else going wrong in the software. Glean, an enterprise AI company with more than a thousand employees, built the agent in-house, and once it was running it handled roughly 95 percent of those issues automatically.
The spending is what made the result uncertain. Glean was paying a million dollars a month for that one agent, a cost that left a real question over whether the agent was more efficient than the humans it relieved.
Glean expects inference costs to come down by orders of magnitude over the coming years, and bills like this one sit inside that bet.
Why does the price of inference keep falling?
Several things push it down at once: cheaper and faster chips, software that runs models more efficiently, and smaller models trained to imitate bigger ones. Epoch AI, a research group, estimated in 2025 that the price of reaching a fixed level of performance fell by anywhere from about 9 times to 900 times per year, depending on the task. Providers also compete on price, which passes those savings along quickly. The result is that a given level of capability costs far less each year than it did the year before.
How does an agent triage a production alert?
It starts by reading the alert, then pulls the related logs, metrics, and recent code changes. Next it compares what it finds against past incidents and written runbooks, which are step-by-step guides for known problems. If the cause is familiar, it can apply the documented fix, and if not, it hands the case to a human with the evidence already gathered. Site reliability teams have kept runbooks since long before AI agents, which is part of why this work is a natural fit for automation.
More from 20VC with Harry Stebbings
