Korva is a scroll of the internet’s best ideas.
What a harness does
A model works like a brain that needs arms, legs, hands, and eyes before it can do work. The harness supplies those, and wrapping a model in one is what makes it agentic. The industry has invented many kinds of harnesses.
Where costs diverge
Pulling apart different models with different harnesses at 8090, a company with its own labs function, shows that they are wildly variant in cost and quality. The models are converging while the harnesses differ enormously, so performance lands all over the map. Buyers then have to decide what they are willing to pay for.
Pressure on model makers
The same shift changes the business of selling tokens. The first version of the AI trade was simple: models differed enough that wrapping tokens and passing them through created value. When models cluster, a company serving a token of diminishing value watches the downstream value get absorbed by the people wrapping its intelligence.
That pushes model makers up the stack. They have to do the work themselves, in areas like cybersecurity, law, and customer support, where it was once unclear whether they would compete.
How does a harness make a model agentic?
A harness runs the model in a loop. The model reads the task and proposes an action, such as reading a file, running a search, or calling a tool. The harness carries out that action, returns the result to the model, and repeats until the task is done. It also decides what the model remembers between steps and what it is permitted to touch.
Why do identical models score differently?
Benchmarks for coding agents, such as SWE-bench from Princeton researchers in 2023, give a model a real software issue and check whether its fix passes the tests. The score depends on more than the model. The prompts, the tools offered, and how the harness manages context all change the outcome. Published leaderboards for the same model often list several different results for that reason.
Why can a cheaper model cost more overall?
A task costs the number of tokens used times the price per token. A weaker model may need more steps, more retries, or longer reasoning before it reaches an answer. In an agent loop, each new step typically resends the earlier conversation as input, so the token count grows with every extra step. A model priced lower per token can therefore finish with a higher total bill.
More from All-In with Chamath, Jason, Sacks & Friedberg
