Korva is a scroll of the internet’s best ideas.
One frontier model, many cheap helpers
Open weight models are generally strong at deterministic tasks: ones where the shape of the output is known, the type of problem is known, and the problem has been solved before. Classifying some text is the standard example. Picture a pile of support messages that need sorting into categories.
The orchestrator calls out to sub-agents whenever such a task comes up. The sub-agents are very low cost and stay focused on that one deterministic job.
The orchestrator then reads the results and goes back to the unknown, non-deterministic task it set out to do. A low-cost model from OpenRouter handles the sorting while the frontier model spends its effort where it is needed.
Why can cheap models handle classification so well?
Classification has a fixed output format and a task that appears constantly in training data, so a small model can learn it without much capacity. Fine-tuned small models have often matched or beaten far larger general models on narrow tasks like this. The cost of running a model rises with its size, so the same job on a small model uses much less compute.
How do an orchestrator and its sub-agents work together?
Anthropic described its multi-agent research system in a 2025 engineering post. A lead agent breaks a question into parts and starts several sub-agents in parallel, each searching its own slice. Each sub-agent returns a condensed summary to the lead, which combines them into one answer.
Why does the output of a task need to be predictable?
When the shape of the output is known, a program can check it automatically, such as confirming that a label is one of a fixed list. Predictable tasks can be tested against examples with known correct answers. Open-ended tasks have no such answer key, which is why they tend to go to the strongest model.
More from 20VC with Harry Stebbings
