Dodona Agent Kernel

Dodona is a startup I've been working on. It's a market where companies lock in the price of GPU hours before they need them, and a lot of those companies already run on AI agents. The founders of Ornn, a similar exchange, told me what happens when those agents trade for more than a few days. Their own history fills up the prompt, and they forget their limits or just stop. So I built what an operating system would do for them. The model is the CPU, and the kernel is everything you shouldn't trust a CPU with. It decides which agent runs next, keeps every prompt inside a fixed budget by setting older history aside and summarizing it with a cheaper model, and checks every order in plain integer arithmetic before it reaches the market. The kernel is written in C++ and talks to the Dodona platform over a simple line protocol, so any assistant that speaks MCP can hand it a goal and get a booked order back without ever seeing a prompt. I measured it against a normal, widely used LangGraph trading agent over thirty simulated days with the same model and the same market. The plain agent's prompt grew from fifteen hundred tokens to sixteen thousand. Under the kernel it stayed flat at around eleven hundred, cost a third as much, and protected the same amount of compute. The kernel also refused seven orders the agent should not have placed, and the benchmark caught a bug in my own code, which is now fixed with a test. The kernel, its tests and benchmarks are in the linked repo. The Dodona side, meaning the driver, the live console and the MCP tools, is on the feature/agent-kernel branch of github.com/pranavrp122/dodona-website. Dodona's order book existed before the hackathon. Everything above was built during it.