Korva is a scroll of the internet’s best ideas.
Whoever runs the agent keeps the learning
A task usually starts with a documented process, maybe ten steps for completing a piece of work. Over time people optimize and tweak it, and much of that never gets written down. It lives in what they learned by doing the work again and again, and they apply it in real time.
If agents do the work, that institutional learning builds up in the agent. A company that does not control the agent running its work, and does not own what the agent learns over the years, ends up fully dependent on the AI company behind it. This goes beyond technology dependence. It is operational dependence on whoever runs those agents.
That puts a question in front of every enterprise: how to use these AI technologies while keeping control of the compounding learnings, which belong to the enterprise. Worry about data being used to train models has mostly faded for companies that sign the right contract, so cost is what currently pushes companies toward open source models.
Why does unwritten know-how matter so much?
The philosopher Michael Polanyi described this in his 1966 book The Tacit Dimension with the line that we know more than we can tell. Skills like judging when a process step can be skipped or spotting a problem early are learned by repetition and rarely make it into a manual. Organizations that lose experienced people often find that written procedures alone cannot rebuild what those people knew. The same logic applies to whatever does the work, whether a person or an agent.
Why is switching AI providers hard?
Over time, an agent's prompts, tools, memory, and tuned workflows get built around one provider's platform. Moving them means rebuilding and retesting each piece. Economists Carl Shapiro and Hal Varian described this pattern in their 1998 book Information Rules, calling the cost of leaving a system a switching cost. The more a company invests in one system, the larger those costs grow.
More from 20VC with Harry Stebbings
