# Glossary
**AgentRun** — The single primitive: Identity + BrainPolicy + AutonomyContract + ToolSet + Memory + Workspace.
**Brain** — A model behind a uniform trait. Normalizes messages, streaming, and tool-calling.
**Router** — Selects the model chain per task and budget. No user lock-in.
**Autonomy dial** — Continuous trust contract (Supervised → Trusted → Autonomous). Not two modes.
**Checkpoint** — Reversible git snapshot created before every mutating batch.
**Swarm** — Multiple AgentRuns over a shared workspace. Default: Planner → Coder → Verifier.
**Skill** — A reusable `SKILL.md` capability loaded into agent context when relevant.
**Curator** — The self-improvement loop: grades, consolidates, deduplicates, prunes skills.
**Surface** — A thin client over the headless runtime (TUI, CLI, API, gateway).
**Transcript** — Append-only event log (`inputs.json` + `events.jsonl`) enabling replay.
**SOUL** — Persistent agent identity file (TOML): name, role, personality, rules, default model.
**MCP** — Model Context Protocol: external tool servers connected via stdio/HTTP/SSE.
**OrgPolicy** — Enterprise policy: autonomy ceiling, provider allow-list, budget caps, blocked paths, SSO.
**Hallucination guard** — Prevents agent from claiming results without real tool execution.
**Anti-simulation guard** — Rejects assistant turns that fabricate test/build/output claims.
**Hook** — Lifecycle event handler (shell or builtin), blocking or non-blocking, 12 events.
**Reasoning depth** — Adaptive planning: trivial tasks get minimal steps, complex tasks get deep decomposition.