Vector retrieval (RAG) as a tool-selection strategy for autonomous agents has officially hit a dead end. Researchers from Fudan University and the National University of Singapore have confirmed that fishing for the right software in massive libraries based solely on "name similarity" leads to structural blindness. When an agent selects a skill based only on embeddings, it ignores functional dependencies and often pulls conflicting tools into the context. At a scale of thousands of entries, this "flat" method becomes a primary point of failure, turning task execution into a logical mess.
SkillDAG proposes replacing primitive vector lists with Typed Skill Graphs (Directed Acyclic Graphs). Unlike the rigid pipelines of the past, SkillDAG opens a structural interface directly to the LLM. In each cycle, the agent receives more than just search results; it gets three distinct fields: matches, neighbors, and conflicts. This forces models—whether MiniMax-M2.7 or gpt-5.2-codex—to consciously analyze why a specific skill was chosen and which auxiliary utilities it requires.
Key Architecture of SkillDAG
The system evolves via a "propose-commit" protocol: the agent registers new connections that are verified for cycles and contradictions before being written to the registry. Monotonic online editing allows for dynamic knowledge base updates without data loss. The model receives a complete dependency map, which eliminates the use of mutually exclusive tools.
It is time to stop treating an agent's toolkit like a junk pile of scripts and start viewing it as a dependency graph.
Data confirms that architectural rigor beats "clever prompting." On ALFWorld and SkillsBench benchmarks, SkillDAG achieved a 67.1% success rate, outperforming the baseline Graph-of-Skills by 12.8 points. More importantly for enterprise applications, the solution remains stable as the skill pool grows tenfold, whereas traditional methods degrade. By utilizing monotonic online editing, SkillDAG boosts recall from 65.5% to 78.2%.
If your autonomous systems are stumbling over software conflicts or forgetting prerequisites, moving to typed skill graphs is no longer optional—it is a matter of project survival in production. Structured retrieval is set to become the standard for any agent managing anything more complex than a handful of specialized functions.