Autonomous Agents for Scientific Tasks - Sina Shahandeh, Radicait
AI coding agents plateau on scientific tasks because they run out of novel hypotheses, not implementation ability.
“they ran out of ideas uh or you know what people call them research taste”
Sina Shahandeh of Radicait argues that AI coding agents are competent at running experiments and implementing code but hit a ceiling on long-horizon scientific tasks because they cannot generate novel research hypotheses—what he calls 'research taste.' The talk frames hypothesis generation, not memory or execution, as the core unsolved bottleneck for autonomous scientific agents. Radicait is applying this framing to in-silico PET image synthesis from CT scans as a concrete use case.