Agents quickstart

First recommendation in an afternoon. An agent in ReasonRank is one production workload you want to right-size: eval suite (test cases) + production model + volume + optional BYO harness. The product loop is: evaluate → recommend → gate in CI → verify dollars.

1. Create an agent

Dashboard → Agents → New agent (blank) or Start from template (fastest).

Name it after the real workload (e.g. support-triage, doc-qa). Add a short description so teammates know what “quality” means.

2. Attach how it runs (BYOK / BYO harness)

Details: /docs/byo-harness.

3. Add ≥5 test cases

Open the agent → test cases. Use exact-match / regex / JSON checks where you can; LLM-as-judge when the answer is free-form. Five shared cases is the minimum for a statistically gated recommendation.

4. Set production model + volume

Pick the expensive baseline as the production model. Volume comes from ingested traces (Settings → Ingestion) or a monthly override so savings project against real traffic. If you're on Vercel AI SDK, OpenAI Agents SDK, LangGraph, or Pydantic-AI, a one-call adapter wires trace export for you — see /docs/adapters; any other stack can point a plain OTel exporter at ingest.

5. Run → recommend → verify

Run candidate models against the suite. Open Savings for the recommendation (cost per success + non-inferiority CI). Apply in your stack or ours, then confirm verified dollars once live traffic reflects the switch.

6. Optional: CI gate

Same math as Savings — fail the PR if quality slips: see /docs/ci-gating. Prefer running on your own machine or CI box? reasonrank run --local executes the suite with your env provider keys and submits outputs for server-side scoring, so local, CI, and cloud runs share one statistical gate.

---

What “done” looks like: one agent, ≥5 cases, a cheaper model that clears the quality gate, and a verified $/mo card you can show finance.