#358 How AI Agents Will Work While You Sleep | Ruslan Salakhutdinov, Professor at Carnegie Mellon

The conversation with Carnegie Mellon professor Ruslan Salakhutdinov highlights how AI agents have moved from experimental demos to practical tools that can write code, fill forms, and browse the web. Recent releases from Anthropic, OpenAI and other frontier labs show agents achieving 45‑50 % success on complex tasks, a level that rivals early human‑in‑the‑loop workflows. By automating routine computer‑use activities, these systems free professionals to focus on higher‑value decisions, while the underlying models benefit from larger training datasets and improved reasoning pipelines. This shift toward autonomy is reshaping productivity across software development, data analysis, and everyday office work.

Despite the progress, extending agent operation to long‑horizon tasks remains a research bottleneck. Tasks that last several hours—or eventually days—require more than a single binary reward; they need intermediate signals that tell the model which steps are correct. Salakhutdinov describes the emerging “rubric‑based” approach, where partial credit is assigned to sub‑tasks, addressing the classic credit‑assignment problem in reinforcement learning. By providing richer feedback, agents can backtrack, retry, and refine their plans, moving beyond simple unit‑test verification toward robust, multi‑step problem solving in domains such as scientific computing and complex web workflows.

The next wave focuses on multi‑agent orchestration, where a powerful planner delegates subtasks to smaller, cheaper models that execute locally. This swarm‑like architecture enables parallel processing, reduces inference costs, and improves scalability for enterprise deployments. For businesses, the payoff is tangible: overnight experiment monitoring, automated job‑search aggregation, and continuous form‑filling without human supervision. As agents become reliable enough to run while employees sleep, organizations can capture hidden productivity, cut compute waste, and accelerate decision cycles. The convergence of better reward structures and coordinated multi‑agent systems signals a near‑term transition toward truly autonomous digital assistants in the workplace.