All docs
ConceptsAnswers, sessions and agentsdashboard

Answers, sessions and agents

What you send: answers, the sessions that group them, and the agents and hand-offs inside a workflow.

Answer (event)

What it is. One question and the answer your AI gave, logged from your app, with whatever context you have: sources, agent, session, tags, metadata, SQL, tools.

Why it matters. It is the unit everything is scored on. Each optional field turns on more checks.

How it relates. On arrival an answer is pinned to the policy version in force and the scoring rule that matched it, then scored in the background.

Where you see it. Evaluation › Live events, and the answer's drawer (Summary, Scores, Claims, Checks, History, Details). See Live events and Event details.

Example.

client.log(
    "What is the in-network deductible on Gold PPO 2026?",
    "The individual in-network deductible is $1,500.",
    reference_materials=[{"id": "sob-2026-p2", "source": "Summary of Benefits 2026, p.2",
                          "content": "Deductible: $1,500 individual."}],
    agent_name="benefits-agent", session_id="chat-4821", environment="prod",
)

Session and steps

What it is. The steps of one conversation or workflow, grouped by the same session_id. Each step is an answer.

Why it matters. Problems often sit between steps: a field changed in a hand-off, or one weak step dragging down the final answer.

How it relates. A session shows the agent path, every hand-off and its weakest step. Sampling keeps a whole session in or out together.

Where you see it. Evaluation › Sessions (views All sessions, Hand-off failed, A step below 0.75) and the session detail. See Sessions.

Example. Claim CLM-20931: intake → adjudication → letter-writer, three steps in one session.

Agents and hand-offs

What it is. An agent is the part of your app that produced a step (agent_name). A hand-off is the work one agent passes to the next: the fields and the message (handoff=).

Why it matters. Hand-off fields are compared as sent and as received, so a date or leave type rewritten between agents is caught and the hand-off is named as the place to correct.

How it relates. Checked against your hand-off contracts. The final step can be marked with is_aggregator=True. LangGraph, Google ADK and MLflow send agents and hand-offs for you.

Where you see it. Agent flow and Hand-offs in a session; the Agent filter in Live events; Group by agent in Issues.

Example. Intake sends onset_date: 2026-03-01; adjudication receives 2026-01-03. Hand-off fields unchanged fails; where to correct: hand-off.