Agent Data
Agents produce two layers of data: everything a human participant would produce, plus decision-level logs of their internal reasoning.
What gets recorded
Human-parity data. Agents emit the same component lifecycle and response events as human participants — component views, ratings, text responses, chat messages. Your existing Events and Components tabs, exports, and analysis code work on agent rooms unchanged.
Decisions. Every agent turn is logged as a decision record: what the agent perceived, the prompt it received, its reasoning, and the action it chose. Agents with cognition enabled also log theory-of-mind signals (prediction error / surprise, regret, per-peer belief snapshots) as analyzable dependent variables.
Run manifests. Each agent's run gets a manifest recording the model and provider used, token and cost totals, the random seed, the agent (persona) ID, and the platform code version — everything needed to reproduce or audit a run.
Viewing agent data
Open Data Management for an experiment with agent runs:
- The Agent Decisions tab lists per-turn decisions, grouped by run, with drill-down from a run manifest into its decisions and from a decision into its full detail (prompt, reasoning chain, belief snapshots).
- Runs flagged as orphaned belong to rooms that ended abnormally — useful for spotting rooms killed by budgets or timeouts.
Clicking through to a participant opens their Events / Reasoning / Variables views, with per-run manifests and CSV/JSON export:
Exporting
Agent decisions are exportable from Data Management like any other data type (export type AGENT_DECISIONS). Standard exports (events, components, chat) include agent participants automatically.
API access
Agent data is available through the HyperStudy REST API with an API key that has the read:events scope:
| Endpoint | Returns |
|---|---|
GET /api/v3/data/agent-decisions/experiment/{experimentId} | All decisions + run manifests across every room (supports ?limit=, ?detail=true) |
GET /api/v3/data/agent-decisions/room/{roomId} | Decisions + run manifest for one room |
GET /api/v3/data/agent-decisions/room/{roomId}/decision/{decisionId} | One decision with full detail blobs |
GET /api/v3/data/agent-runs/experiment/{experimentId} | All run manifests, with orphaned-run flags |
The Python SDK wraps these as get_agent_decisions(), get_agent_decision(), and get_agent_runs().
Some recorded fields (cognition configuration snapshots, fine-grained theory-of-mind traces) are still being surfaced in exports — the underlying records are complete, but export coverage is evolving alongside the cognition system.