Core concepts for the platform: agents, orchestration, knowledge, tools, memory, policies, observability, and evaluation.

Concepts

Overview

Concept pages define the vocabulary used across the platform. They describe responsibilities, boundaries, and the reasons certain controls exist (security, reliability, and audit).

What this documentation covers

  • Agents and tool-calling contracts.
  • Orchestration and state machines (checkpointing, replay).
  • Knowledge layer and retrieval (hybrid search, reranking, citations).
  • Policies and permissions (RBAC/ABAC, approvals).
  • Memory boundaries and retention.
  • Observability, quality signals, and cost attribution.
  • Evaluation harnesses, golden sets, and red teaming.

Who this is for

  • Readers who need shared terminology across engineering, security, and product.

Operational expectations

  • Examples favor deterministic behavior and explicit contracts (schemas, IDs, correlation).
  • Security is treated as a default: least privilege, audit logs, and safe fallbacks.
  • Changes are versioned; breaking changes should be announced via /docs/changelog.