Research notes: tool and capability graphs

This Advanced note supports agent grounding and tool recall. It is for application developers. A student using an existing agent should start with Work with an agent.

What the graph represents

Tools, declared capabilities, inputs, outputs, dependencies, and deprecation can be stored as ordinary graph records. Cypher retrieves declared relationships; forge.find retrieves candidates from names/descriptions or a configured vector space. Neither executes tools nor enforces hosted permissions.

Share reference nodes under stable keys so a dependency refers to the intended input or capability. A node labeled Role or an edge named MAY_USE records application metadata; actual authorization remains at the execution boundary.

What requires evaluation

Test capability lookup, missing inputs, deprecated tools, ambiguous intent, and unauthorized actions on the intended registry. Measure candidate recall and incorrect selections separately from latency. Do not infer agent reasoning quality from a successful graph query or from the absence of a network hop.

Native examples are linked from the worked guide. Its synthetic notebook checks exercise bounded behavior. They do not prove real-agent quality, a branded integration, or universal sub-millisecond lookup. For durable registries and transfer, use save and reopen and portable projects.