Skip to main content
Agent dispatch fires a hosted coding agent when a Latitude signal needs engineering follow-up. Latitude sends the agent a context-rich prompt — the signal, a link back to Latitude, and a few sample traces — and the agent reads Latitude back over the MCP connection to root-cause the issue and implement a fix.

​
How it works

  1. A signal escalates (or an incident opens) on a monitored project.
  2. Latitude assembles a prompt with the signalId, a deep link, and sample trace ids, and fires the configured dispatch target.
  3. The agent — already connected to the Latitude MCP — investigates the signal (below), implements a fix, runs the project’s checks, and reports what changed.
Latitude is the trigger and context provider, not the agent runtime: it wakes the agent and hands it the data, but does not run the fix or wait on a completion loop.

​
Supported targets

Configure a target per project under Settings → Integrations:

​
Investigate a signal

Once woken, the agent walks the signal end to end over MCP — the curated tools for orientation, then queryAnalytics and querySpans for the questions no single endpoint answers. 1. Orient with the signal tools. These encode Latitude’s semantics (baselines, escalation, co-occurrence) so the agent doesn’t reconstruct them. 
getSignal({ projectSlug, signalSlug })         // lifecycle, source, priority, evaluation
getSignalTrend({ projectSlug, signalSlug })    // occurrence trend — the escalation shape
listSignalTraces({ projectSlug, signalSlug })  // the member traces
getTrace({ projectSlug, traceId })             // → listTraceSpans → getTraceSpan to read the failure
2. Slice it — “is this concentrated in one model?” A signal’s occurrences are scores carrying its signalId, so the scores stream of queryAnalytics answers breakdowns that have no dedicated endpoint: 
queryAnalytics({
  stream: "scores",
  filters: { "score.signalId": [{ op: "eq", value: "sig_9f2…" }] },
  metric: { kind: "count" },
  breakdown: "model",
  range: { fromIso: "2026-06-23T00:00:00Z", toIso: "2026-06-30T00:00:00Z" },
  orderBy: { by: "value", direction: "desc" }
})
// → [ { key: "gpt-4o-mini", value: 412 }, { key: "gpt-4o", value: 38 }, … ]
90% of occurrences on one model points at a model-specific root cause. Switch stream, metric, or breakdown to confirm impact from another angle (e.g. stream: "traces", metric: "errorRate", timeBucket: { unit: "week" }). Values come back in human units — duration in seconds, cost in dollars, rates as a 0–1 fraction. 3. Drill to the evidence. queryAnalytics says where the problem is; querySpans hands you the actual spans to read — including failures buried inside traces that otherwise succeeded (which a trace-level error filter would miss): 
querySpans({
  filters: { toolName: [{ op: "eq", value: "search_docs" }], operation: [{ op: "eq", value: "execute_tool" }] },
  range: { fromIso: "2026-06-23T00:00:00Z", toIso: "2026-06-30T00:00:00Z" },
  limit: 20
})
// → { items: [ …the individual failing search_docs spans with their args + errors… ], hasMore: false }
The agent reads the spans, identifies the pattern, and implements the fix — closing the loop the dispatch opened.
The same query tools power dashboards. See Build a dashboard from Latitude data to turn these queries into a self-contained HTML report.