Skip to content

Core Concepts

An Agent packages:

  • a name
  • instructions
  • a model identifier
  • zero or more tools
  • zero or more handoffs
  • optional output guardrails

A tool is callable code exposed to the model. In aioc, tool execution is never a direct consequence of model output. Tool calls are proposals that still need deterministic authorization.

A handoff lets one agent delegate to another. Internally, aioc models handoffs as runtime-managed tool proposals so they can pass through the same governance boundary as normal tools.

Policies are deterministic functions that decide whether a proposed action may proceed. Current stable behavior is default deny unless a policy returns allow.

RunRecord is the run-level audit artifact emitted by the runtime when recording is enabled. It captures:

  • input/output trajectory items
  • prompt snapshots
  • request fingerprints
  • policy decisions
  • optional guardrail decisions
  • metadata and redacted context snapshot