Documentation Index
Fetch the complete documentation index at: https://docs.latitude.so/llms.txt
Use this file to discover all available pages before exploring further.
Latitude Overview
Latitude is an open-source, MIT-licensed platform for improving production AI agents. It helps teams capture real agent traffic, find important behaviours across that traffic, and turn repeated failures into trackable issues. The core workflow is simple: observe what happened, search for what matters, track recurring issues, and monitor whether fixes work.
Observability
Latitude starts by collecting agent telemetry. Each interaction becomes a trace made of spans for LLM calls, tool calls, retrieval steps, HTTP requests, and other instrumented work. Related traces can be grouped into sessions so teams can review full multi-turn conversations. Use observability to inspect:- traces, spans, and sessions
- model providers, model names, tokens, latency, cost, and errors
- tool calls and nested execution paths
- users, session ids, tags, metadata, environments, and releases
Search
Search lets teams find behaviours across production traces without knowing the exact words users used. Latitude supports semantic search, exact text search, and metadata filters, so you can move from a broad question to a focused set of real examples. Use search to find cohorts such as:- frustrated users
- tool loops or tool call failures
- incorrect answers or hallucinations
- failed purchases, onboarding problems, or domain-specific workflows
- traces from a specific user, model, release, environment, tag, or metadata value
Issues
Issues are recurring failure patterns discovered from failed signals. Human annotations, built-in flaggers, evaluations, and custom scores can all identify bad behaviour. Latitude groups similar failures into issues with examples, trends, affected users, lifecycle states, and linked monitors. Use issues to:- triage new and escalating production problems
- inspect example traces that explain the failure pattern
- distinguish noise from problems worth fixing
- resolve or ignore issues as part of your team workflow
- generate evaluations that keep monitoring for the same failure mode
Why Latitude?
AI agents fail in ways that are hard to predict upfront. Logs and dashboards show individual events, but teams also need to understand repeated behavioural failures: what users experienced, how often it happened, who was affected, and whether a fix prevented it from coming back. Latitude is built around that loop:- Agent-native telemetry: inspect multi-step agent behaviour, not just isolated API calls.
- Semantic discovery: search by meaning across production conversations.
- Issue-centric workflow: turn failures into named, trackable problems instead of scattered examples.
- Human-aligned monitoring: use annotations, evaluations, and scores together so automated monitoring stays connected to human judgment.
- Open source and MIT licensed: run, inspect, self-host, fork, and contribute to the platform your team depends on.
How Latitude is different
Most LLM observability platforms focus on one narrow part of the production quality loop. Tools such as Langfuse, Datadog, Sentry, and PostHog are useful for visibility, evaluation, analytics, or feedback workflows, but they usually leave the work of turning scattered signals into trackable AI-agent issues to the team using them. Latitude is built around LLM issue discovery. Observability, semantic search, annotations, flaggers, scores, and evaluations are designed to work together automatically so teams can track production agent failures with minimal manual setup. The difference is the combined workflow:- Flaggers detect common failure categories automatically, such as frustration, refusal, jailbreaking, tool errors, and empty responses.
- Semantic search helps teams find product-specific behaviours and failure modes across real conversations.
- Annotations turn human review into structured signal that can create or refine issues.
- Automatically human-aligned evaluations convert important issue patterns into monitors that stay connected to human judgment.
- Issues bring those signals together into named, prioritized, lifecycle-managed production failure patterns.
Getting started
- Start tracing: connect your agent and send your first traces to Latitude.
- How to use Latitude: follow the recommended workflow for finding, tracking, and fixing production issues.