The procedural gap is the name A Comprehensive Survey on Agent Skills: Taxonomy, Techniques, and Applications gives to the distance between an LLM-based agent’s access to external capabilities and robust execution of a task with them. Access to tools — through APIs, plugins or protocol layers such as the Model Context Protocol — does not determine when a capability should be invoked, how multiple tools should be coordinated, how failures should be handled, or how outputs should be validated. When that orchestration is left to the model at inference time, the survey argues, it becomes a major source of brittleness, high latency and unreliability, particularly as tasks become long-horizon and heterogeneous.
Usage
The survey uses the term to locate Agent Skills in an agent’s capability stack. It distinguishes passive knowledge absorbed into a model’s parameters through pre-training, fine-tuning and alignment — static, opaque and often weak in specialized or fast-changing domains — from active knowledge obtained at runtime by retrieving documents, invoking tools and APIs, accessing MCP servers and executing skills. Active knowledge is more dynamic and grounded, but on the survey’s account access to it still does not say what should be used, when, in what sequence, or how its outputs should be checked; that is the gap.
The survey’s illustration is that a tool exposes an atomic capability — what can be done, not how it should be used. A search tool does not say when search is preferable to memory retrieval; an API tool does not say what to do when a schema changes; a code interpreter does not say how outputs should be validated. It reads MCP and similar infrastructure as solving an interoperability problem rather than this procedural one, and presents skills — reusable procedural artifacts specifying when and how external capabilities should be applied — as what closes it.
Related Terms
- Agent Skills — what the survey proposes as the bridge across the gap
- Function calling — the capability access the gap sits on top of
- Model Context Protocol — infrastructure the survey reads as solving interoperability, not this gap