Auker 1.0: Self-improving agents that spend intelligence better

Auker is Reasonary AI’s self-improving agentic model. It breaks complex tasks into subproblems and learns how to allocate models, agents, tools, and reasoning to each one. In our MMMU-Pro evaluation, 116 optimization experiments produced a policy that reached 80.35% accuracy at a total cost of $9.73—about 6× lower cost than Auker’s initial baseline at essentially unchanged accuracy. We’re bringing this adaptive allocation approach to enterprise workloads.

Hi Product Hunt — we’re excited to introduce Auker 1.0. We asked whether an agent could improve the way it works, not only its answers. Auker learns a policy for task decomposition, model routing, agent coordination, tool use, and reasoning. After 116 optimization experiments on MMMU-Pro, it reached 80.35% accuracy at $9.73—about 6× below its initial baseline cost at essentially unchanged accuracy. Within our tested setup, it also moved the cost–accuracy frontier beyond GPT-5.6 Luna and Qwen3.7-Plus at 25–30% lower cost. We’re now testing this approach on real enterprise workloads. What workload would you test first, and what evidence would make you trust adaptive allocation? Technical report: https://reasonary.ai/blog/auker-...