New AI System Plans Complex Research Projects Autonomously.

Autonomous research agents are becoming increasingly capable, handling tasks from literature search to manuscript drafting. However, managing an entire research project, which involves multiple interconnected tasks, parallel alternatives, and complex dependencies, remains a significant challenge. Existing systems often treat a project as a single, oversized task or produce vague, overlapping task lists, leaving critical coordination to manual effort. To address this, researchers introduce Project2Task, a novel graph-guided planning layer designed for autonomous research at the project level. Given a high-level project brief, Project2Task first identifies potential contributions, termed "innovation atoms," and structures them within a directed lineage graph. This graph allows for strategic selection among horizontal, vertical, or hybrid portfolio decompositions using a lightweight Bernoulli block-model objective. The system then generates well-defined, bounded tasks with clear ownership, resolves overlaps, fills in missing execution details, and outputs "task contracts." These contracts explicitly specify objectives, required inputs, expected artifacts, evaluation criteria, boundary constraints, dependencies, and execution order. Crucially, these contracts are executor-agnostic, facilitating integration of task outputs into a cohesive project-level result. Evaluations show Project2Task significantly improves project quality and downstream task accuracy compared to baseline methods, demonstrating the value of explicit project-to-task planning for coherent and executable research portfolios.