Korva is a scroll of the internet’s best ideas.
More agents mean more exceptions
A sales agent has a margin threshold it can't cross when promising a client a price. Curative doesn't want it to simply say no, so it sends the case up for a business decision about whether the deal is right for that client. If an approval request happens only 1% of the time but the agent does 10 times as much work, that is 10 times as many requests as last year.
Curative tried agents supervising agents, which works to a degree. A more expensive model could supervise even where it would cost too much to run the core workflow. The hard part is that exceptions need a nuanced decision, not a rubber stamp of yes or no either way.
Agent supervisor is the role Curative expects to be big in five years and doesn't exist today. The best preparation is to play with the models. Many people underestimate what they can do, perhaps because they tried ChatGPT two years ago, and the models are much better than they were six months ago.
Why do exceptions limit automation gains?
Amdahl's law, stated by computer architect Gene Amdahl in 1967, says the overall speedup from improving part of a process is limited by the part left unimproved. If agents handle routine cases and people handle exceptions, human review becomes the limit. Faster agents then just move the queue to the approval step. The same pattern appears in factories and call centers.
How does one model supervise another?
A stronger or more expensive model can review the work of a cheaper one, a practice often called LLM-as-judge. Checking an answer is usually easier than producing it, so the reviewer can be used on a sample or only on flagged cases. Researchers have found such judges agree well with human raters on many tasks but have biases of their own, as shown in a 2023 paper on MT-Bench. Systems often combine automated review with human sign-off for the hardest calls.
Why are exceptions hard to approve?
Exceptions are the cases that break standard rules, so each one needs context that a rule cannot supply. A pricing request below a margin threshold may be worth accepting for a long-term client. Reviewers weigh the account's history, its future potential, and the cost of refusing. Those factors vary by case, which makes them resistant to simple yes-or-no automation.
More from 20VC with Harry Stebbings
