Introduction

Morse AI provides verdict-first performance intelligence for production agents: know which agents are getting better — and which are just getting expensive. The familiar category is AI-agent observability and evals. The operating model is performance brief first, traces on demand.

A verdict is a ranked performance brief, not a fused cost × eval metric or a root-cause claim. It keeps labeled or confirmed production outcomes, cost per successful outcome (CPSO), and production success trends separate from applicable quality signals computed on eligible sampled traces.

The workflow is measure → prioritize → verify:

1. Your team maps and confirms what production success means.
2. Morse performs a first-pass review and ranks the agents and cohorts that deserve attention.
3. You verify the result against supporting cohorts, traces, and signals before acting.

What “correlated in a single trace” means

A trace is an ordered record of the LLM calls, tool calls, database queries, and HTTP requests captured for a run. Spans are typed from the start — agent, llm, tool, http, db — so agent and available infrastructure signals can be reviewed in the same waterfall. Infra scope is deliberately narrow: trace-anchored correlation only, not fleet, cluster, or K8s monitoring.

What you get

• Performance briefs — ranked agents and cohorts, with the evidence lanes kept separate and unsupported conclusions identified. See the Dashboard.
• Traces — the full run, span by span, filterable down to the cohort you check daily. See Traces.
• Agents — the same runs, rolled up per agent. See Agents.
• Cost — spend and CPSO from labeled or confirmed outcomes. Pricing coverage is the share of pricing-eligible spans Morse successfully priced. See Cost.
• Context — context-window and memory profiling. Likely-producer attribution is best effort, memory comparisons are directional, and prompt-cache savings are estimates. See Context.
• Quality — applicable default scorers evaluate eligible sampled traces; these signals stay separate from production success. Failures, low scores, and novel paths can add candidates to the review queue.
• Reliability — error rates and failure patterns across runs. See Reliability.
• Copilot — ask questions about a trace or cohort from the telemetry Morse has available. See Copilot.

When revision metadata is present, Morse can align outcome, cost, and sampled quality evidence to that revision for comparison. This alignment does not establish that a deploy caused the movement.

Where to go next

• New to Morse AI: start with the Quickstart — instrument one agent, verify a trace, and get to your dashboard.
• Want the vocabulary first: see Concepts for trace, span, thread, and the typed-span model.
• Already emitting OpenTelemetry: see OTLP ingestion — HTTP exporter only today.