Core Concepts

Registered Agent

An AI agent must be registered before it can create Runtime Sessions or submit Agent Runs. Registration answers: which agent is connecting, what tasks is it allowed to perform, and is it currently trusted to run?

• Lifecycle: draft → active → suspended/deprecated. Only active agents may run (a draft agent may opt in for testing via allow_draft_agent when creating a session — never for external Agent Run submissions).
• deprecated is terminal — register a new agent_id (new version) instead of reactivating.
• agent_id is unique within a tenant, not globally — see Security for what tenant scoping does and doesn't mean today.

See Agent Registry for the full API.

Domain

A domain pack (for example healthcare or lending) registers the roles, task types, synthetic cases, reference agents, and policies for one kind of consequential decision. Every agent belongs to one domain, and a request's role must be one the task type allows. The core platform makes no assumption about which domain it is running in.

Runtime Session

One controlled agent workflow. Long-lived: it's the object you poll and act on while work is happening (pause, resume, cancel, approve, reject, retry). A session owns one or more Agent Runs over its lifetime.

See Runtime Sessions.

Agent Run

One execution of an agent within a session. The first attempt is Agent Run #1; a retry creates a new Agent Run, never mutating the previous one. A run's execution_origin is internal (the engine's own reference adapter) or external (a third-party agent submitted via POST /runtime/agent-runs).

See External Agent Runs.

Decision Context Record

The immutable audit record produced by exactly one Agent Run: input hash, retrieved sources, agent output, evidence results, policy results, risk score, and the human review outcome if any. A retried run creates a brand-new record — the previous one is never edited.

See Decision Records.

Decision Controller

The component that turns evidence + policy + risk results into one control action: allow, pause, require_human_review, block, retry, or finalize. It is entirely deterministic — the same inputs always produce the same decision.

Policy Evaluation

Configurable, data-defined rules evaluated against every agent output. Each policy has a condition_type with condition_params (e.g. recommendation_category_present with { category: "decline" }, or claim_type_unsupported with { claim_type: "diagnosis" }), an action (allow, flag, pause, require_human_review, block), and a severity. A policy is either core (domain: null, applied to every run) or scoped to one domain. Policies can be listed, enabled, and disabled at runtime, with no redeploy required. See GET /runtime/policies.

Evidence Evaluation

A deterministic, set-membership check: a claim is "supported" only if every source id it cites resolves to a real retrieved source for that case. This is not NLP-based fact verification — agents must emit structured claims with explicit source citations; there is no claim-extraction step inside the engine.

Runtime Risk

A computed score (0–100, mapped to low/medium/high) derived deterministically from evidence and policy results — never a black-box model score.

Human Review

When the Decision Controller selects require_human_review (or an output is blocked pending review), the session enters waiting_for_human_review. A reviewer then approves, rejects, or triggers a retry via the session action endpoints. See the Human review flow integration pattern.

Decision Replay

Reconstructs the full chronological sequence of a decision (or an entire multi-run session) from its recorded events and Decision Context Record(s) — what happened, in what order, and what the system knew at each step. See Decision Replay.