Something changed this year and not everyone has noticed. Long autonomous runs now produce results at research level: well-known open problems getting resolved with substantial machine assistance, and labs publishing the orchestration prompts behind their proof search rather than only the headline result. The common impression still comes from a chat window: paste a question, get something proof-shaped back, conclude that a few failures settle the matter. A coding agent with a filesystem, a LaTeX compiler, SAGE, Gurobi, and permission to run for hours is a different object.

 

The talk covers what these agents can do on research problems today, how to set one loose on a conjecture of your own, what makes a candidate proof worth trusting, what the whole thing costs in time, money, and attention, and the general techniques for getting started — enough that you can then shape a setup of your own around how you actually work.

许超，电子科技大学计算机科学与工程学院教授，算法与逻辑研究团队成员。主持国家自然科学基金优秀青年科学基金项目（海外）。主要从事理论计算机领域尤其是组合优化的研究。在理论计算机顶会和运筹学顶刊SODA、Mathematical Programming、SIAM Journal on Computing发表论文多篇。