Skip to main content
An Oxygen agent is context-first: what it can do is determined by the context you give it. Each entry in the context: block is a glob (relative to the .agentic.yml file) that pulls in a set of files, and each kind of file grants the agent a specific affordance — a capability it unlocks while reasoning.
The context: block is a list of glob patterns, not named objects — the agent classifies each matched file by its extension and wires up the matching affordance automatically.

​
The affordances

​
Semantic model — the primary affordance

Pointing an agent at your semantic model is the highest-leverage context you can add. With it, the agent:
  • Resolves by name. It searches your catalog for the measures and dimensions a question needs, instead of guessing at raw columns.
  • Compiles deterministically. When the request maps cleanly onto defined measures, the semantic engine compiles the SQL directly — the SQL-generation stage is skipped, and joins and fan-out are handled for you.
  • Stays grounded. When a semantic model is present, the agent’s raw list_tables / describe_table tools are withheld so it doesn’t drift from your definitions.
This is why an agent is only as good as the World Model beneath it: richer semantic context means more questions the agent can answer correctly and deterministically.

​
Example vs. verified queries

Both come from *.sql files, but they grant different affordances:
  • Example queries shape how the agent writes SQL — they’re shown as reference so generated SQL matches your house style. Add a comment at the top of each file explaining what it’s for, so the model has that context.
  • Verified queries are trusted answers. When a question matches one, the agent runs it exactly as written and skips generation entirely — the answer carries a Verified badge. Use these for business-critical questions that must return the same query every time.

​
Domain docs

Plain-Markdown files under your context: globs are injected as domain context, so the agent inherits your business vocabulary — what “active customer” means, how a fiscal quarter is defined, which segment names are canonical. This mostly shapes the Clarify and Interpret stages, where the agent is mapping human language to metrics and back.

​
Automations

Point an agent at *.automation.yml files and it can discover them and delegate a matching request to a pre-built automation instead of solving from scratch — useful when a trusted multi-step pipeline already answers a class of questions.

​
Next steps

How the pipeline works

See which stage each affordance affects

Build a semantic model

The context that matters most