Histeeria sits between your agent and your users as an observability and judgment layer. You send decisions; Histeeria monitors, scores, alerts, and reports — without slowing your agent down.

​
The loop

​
1. Connect

You integrate once:
  • Create a workspace (automatic on signup)
  • Define an agent profile — name, role, and context
  • Generate an API key linked to that profile
  • Add the SDK or call the REST API after each agent turn
Decisions include input (prompt/messages), output (response), optional metadata, tokens, and identifiers (agent_id, session_id, domain).

​
2. Evaluate

The evaluation engine scores each decision on eight judgment dimensions — ethical recognition, uncertainty handling, escalation judgment, and more. Scores are evidence-backed, not binary pass/fail. Evaluation runs asynchronously after ingest. Agents warm up after a minimum number of decisions; then batch evaluation and periodic reports kick in.

​
3. Improve

Use the app to act on results:
  • Command center — live metrics and agent health
  • Monitoring — decision stream and filters
  • Evaluation — dimension breakdowns and trends
  • Reports — periodic judgment summaries
  • Inbox — incidents and items needing review
  • Public profiles — shareable agent judgment pages

​
Design principles

SDK sends are async and fire-and-forget. If Histeeria is unreachable, your agent continues normally — failures are silent by design.
API keys belong to your workspace. All decisions and evaluations are isolated to your organization.
Agent profiles give the judge model context about what this agent is supposed to do, improving score relevance.
Each dimension score is tied to reasoning about the specific decision — useful for audits and debugging.

​
Architecture at a glance

ComponentRole
app.histeeria.comWeb UI — dashboards, profiles, settings
api.histeeria.comREST API — ingest, auth, evaluation data
SDKThin client — queue, batch, retry
Evaluation pipelineLLM judge + aggregation + incident detection

​
Next