Skills configure model-driven nodes
An Agent Skill is not a new Network node type. It is a reusable procedure that can be attached to model-driven nodes:
- LLM Agent — perform a specialised task using an assigned procedure.
- Aggregator — apply a consistent method while combining several inputs.
- Router — use procedural guidance when selecting or shaping a route.
Inherit from a source agent
When a node references an Agent Hub agent, it can inherit that agent’s Skills. This keeps the agent’s tested operating method intact when it participates in a larger workflow.
Set a node-specific Skill selection when the Network needs a narrower or different procedure. The node override takes precedence over inherited bindings.
Choose a context mode
| Mode | Behaviour | Use when |
|---|---|---|
auto | Make Skills discoverable and load relevant instructions on demand | The model and provider support tool-initiated loading |
listing | Expose compact names and descriptions | Another layer will control activation |
inline | Place selected instructions directly in context | Deterministic availability is more important than context cost |
Provider capabilities differ. Test the selected mode with the model used in production.
Tool permissions still apply
The node or source agent must have every tool required by the procedure. Resolving a Skill never grants a capability or connector permission.
Skills explain how; policy decides whether. Use Network policies, tool assignments, connector authorisation, and Human Approval nodes for enforceable control.
Handle unresolved Skills
If a configured Skill cannot be resolved, the node reports it and continues according to the Network’s error behaviour. Monitor unresolved names because they can indicate deletion, version drift, or caller-access problems.
Inspect run telemetry
Network results can include the same activation information used in Agent Hub:
- Available and activated Skills.
- Bundled files read.
- Listing and instruction context loaded.
- Characters added to model context.
- Unresolved Skill references.
Use this alongside node traces and tool events to understand whether a failure came from selection, procedure, permissions, or execution.
Example pattern
For a governed vendor decision:
- An LLM Agent uses Web Research for current external evidence.
- A second node uses Connected Knowledge Search for internal requirements.
- An Aggregator applies Option Comparison to the combined evidence.
- A Human Approval node asks the accountable owner to accept the recommendation.
- A final node uses File Deliverables to create the decision record.
See building workflows and runs and monitoring for the surrounding Network configuration.