Korva is a scroll of the internet’s best ideas.
Let an agent pick its own models
Fixing a customer bug
Picture a customer who reports a bug. The agent investigates what the report says, spins up the product locally, clicks through the same steps to reproduce it, reads the logs, and pinpoints the files involved. Then it debugs, tests again, and puts the fix up for review.
Matching model to step
Different models suit different parts of that work. The hard problems call for maximum thinking and the best model available. Boilerplate and repetitive steps call for an immediate answer that can be verified, as cheaply and quickly as possible.
No loyalty to one lab
Devin is built as a compound model system for that reason. It draws on models from Anthropic, OpenAI, Google, its own models, and open source ones, and chooses dynamically for each task or subtask. Because Cognition isn't tied to one lab, it gains nothing when customers spend more on models. Its incentive is the same as theirs: make the token spend efficient and get real output.
Why do cheap and fast models matter?
AI models are priced per token, and prices vary by more than an order of magnitude between a lab's smallest and largest models. A small model also responds faster, which matters when an agent takes dozens of steps in a row. For simple steps like formatting or renaming, the larger model's extra reasoning adds cost without adding accuracy. Routing each step to the smallest model that can do it lowers both cost and waiting time.
What is a compound AI system?
The term was popularized in a 2024 blog post from Berkeley Artificial Intelligence Research. It describes products built from several components, including multiple models, retrieval, tools, and checking steps. State-of-the-art results increasingly come from the design of the whole system rather than from a single model. A router that sends each request to a different model is one common component.
How does an agent reproduce a bug?
It starts by reading the report and then builds and launches the product in a local environment. It follows the steps the customer described and watches for the same failure. Logs and error traces then point to the code involved. A fix can be checked by repeating the same steps and confirming the failure no longer appears.
More from David Senra
