Researchers have posted a new preprint titled 'Policy Search in Code Space for Multi-Agent Pathfinding' on arXiv. The abstract argues that current approaches to multi-agent coordination rely on either labor-intensive algorithm design or opaque, task-specific reinforcement learning. The authors propose a third route: using autonomous coding agents to search for policies expressed directly as code.
Because the abstract is cut off after 'we show that…', the available record does not include details of the search mechanism or any experimental outcomes. The contribution, as stated, is the framing of multi-agent pathfinding as a code-space search problem enabled by recent progress in coding agents.