Morse Docs
Know which agents are getting better — and which are just getting expensive. Morse ranks your agents in a weekly performance brief by which one needs attention first. Outcomes, cost, and sampled quality stay separate lanes and are never combined into one score; thin evidence is withheld rather than guessed; and the trace behind every line is one click away.
New Morse TV: Morse in three clips
Morse TVListen · 3 short clips · 1:00
- 1What does Morse tell me?
- A weekly brief ranks what needs attention
- Each agent judged against its own normal
- The trace behind every line, one click away
Transcript
Morse tells you which agents are getting better, and which are just getting expensive. Every week, a performance brief ranks what needs your attention first. Each agent is judged against its own normal, not against a fixed bar or other agents. And the trace behind every line is one click away.
- 2Can I trust Morse's numbers?
- Outcomes, cost and quality stay separate
- Never blended into one score
- Thin evidence is withheld, not guessed
Transcript
Morse keeps outcomes, cost, and sampled quality in separate lanes, and never blends them into one score. When the evidence is too thin to judge, the brief says so, and tells you what it still needs, instead of guessing.
- 3What do I need to start?
- Cost and traces from your first run
- Tell Morse what success means
- Verdict: 20 labeled runs per agent in 7 days, 4 days
Transcript
One agent is enough to start. You see cost and traces from your first run. Then tell Morse what success means. A verdict needs twenty labeled runs for an agent in the last seven days, on at least four different days. Until then, Morse tells you it can't judge yet.
Every Morse TV episode is listed on Morse TV.
Start here
- Introduction — what Morse is and how the performance brief works.
- Quickstart — install the SDK and send your first trace.
- Concepts — the mental models: the data model, honest numbers, the account lifecycle, and environments.
Find your way around
- Product Guide — a page per screen, in the same order as the app’s sidebar: Home, Traces, Agents, Reliability, Investigations, Cost & Budget, Efficiency, Quality, Verify, and Settings, with each entry’s tabs listed under it.
- SDK — Python and TypeScript, with adapters for Anthropic, OpenAI Agents, LangGraph, LangChain, the Claude Agent SDK, and the Vercel AI SDK. It fails silent.
- Integrations — already emit OpenTelemetry traces? Send them over OTLP/HTTP, no SDK needed. Or point an MCP client at your own data.
New here? The Quickstart is the fastest path to your first trace.