All docs
Docs
Core Concepts

Agent contexts

Agent contexts define the read model an agent is allowed to use.

Overview

A context is a named bundle of bindings and data sources.

Context names, table names, aliases, output slot names, and search step names are developer-defined. SESSION values such as tenant_id and current_ticket_id are trusted values your backend sets before the context is used.

SEARCH ... QUERY ARG question means the query text is supplied later by the capability caller through INVOKE/PROPOSE JSON or SDK arguments.

The capability references a context instead of letting prompts assemble raw database access.

Developer notes

  • Use one context per trust boundary.
  • Keep tenant, principal, and entitlement values hidden.
  • Version contexts before changing semantics used by customer workflows.
  • Add parser tests before making public grammar changes.