Public reference
Agent Grounding
Definition
Agent grounding is the practice of anchoring AI agent inference and execution in governed, freshness-scored organizational truth rather than ambiguous or stale documentation. Infrastructure and knowledge repositories are treated as organizational cognition substrates: documentation quality directly impacts execution quality.
Key properties
- Canonical truth hierarchy (highest to lowest precedence): runtime state, current platform state, active architecture docs, operational runbooks, historical docs, scratch/drift notes
- AI grounding segmentation: safe grounding sources (canonical ontology, approved workflows, runtime inventories, reviewed architecture, governed operational procedures); contextual grounding sources (planning docs, brainstorming, proposed architecture — used with confidence penalties + provenance tracking); unsafe grounding sources (stale drafts, superseded docs, orphaned notes, experimental scratchpads — excluded from default ingestion)
- Gold vs Blue cognitive separation: gold = experimentation/hypotheses/incubation; blue = governed/deterministic/operationally authoritative, safe for agent grounding