Aster grew out of years spent building AI agents for customers. Every new model brought new possibilities—and another round of evaluations, prompt changes and context tuning.
A lower token price wasn’t always a lower cost to finish the task.
We had agents doing real work for customers. Frontier models gave us a strong starting point, but using them for every step was often slow and expensive. Open-source models gave us more options. The question was whether they could handle our actual workload.
Prompts, tools and conversation history had been tuned around the previous model. A replacement could do well in a benchmark and still miss a step our workflow depended on. So we built datasets, ran evaluations and adjusted the harness. Then the next model arrived.
Some agent tools made one model family the easy default. Others let us connect almost anything. Either way, we still had to decide which model could do the next task well.
As we spoke with more customers, demand kept growing for something that works out of the box. Teams wanted model selection and context management handled for them, whether they were using an existing agent harness or building their own agent workflows. That became part of what we set out to build with Aster.
Test the tasks that matter.
Rework prompts, tools and context.
Look for regressions, not just wins.
The auto-routers we tried didn’t work reliably enough for our agents. They could treat a short follow-up as a simple task and choose a model that lost track of the work. We still had to step in and manage model choice ourselves.
The work behind it
A short message can carry a complicated task.
We needed a router that considered the task, its history and what success looked like. Price and prompt length could not tell us that on their own.
We develop Aster Router through systematic research into reinforcement learning, context-aware selection and caching. Our team builds and evaluates these algorithms in-house, with the full agent workflow in mind.
We research routing as a sequential decision problem: which model should take the next step, given the task and everything that came before? Our reinforcement learning work uses outcome feedback to improve the policy, with quality, cost and latency as its objectives.
Our context-selection research focuses on the instructions, history, tool results and dependencies each task needs. We develop methods to leave irrelevant material behind while preserving the information the next model needs to continue the work.
We’re also developing our own caching algorithms to reuse stable instructions, tool definitions and shared context. This research accounts for prompt-prefix changes, cache expiry and provider and model boundaries, so switching models can be evaluated alongside the context already cached.
Work that helps explain the challenges behind model selection and reliable agents.
We’re the team at VectorQ Lab. Aster is our response to the problems we kept meeting in customer workflows—and the system we wanted for our own agents.
