Latitude turns AI agent chats into an observability signal

Why it matters

As companies put AI agents in front of customers, the next infrastructure fight is over who owns the feedback loop from messy conversations to fixes, evals and retention signals.

Cesar Miguelanez (@heycesr) put Latitude in front of a sharper agent-observability problem Tuesday: production AI agents are talking to customers at scale, but most of what those conversations reveal about product failure still dies inside chat logs.

https://x.com/heycesr/status/2069435941358325774

In a 36-post thread on X, Miguelanez argued that an AI agent's conversations may be the most underused dataset inside a company. His claim is simple and pointed at teams that have already moved agents into customer-facing workflows: latency, cost, traces and dashboards can show that something happened, but the useful signal is often in what the user asked, where the agent misunderstood, when the user got frustrated, and what repeated failures say about churn risk.

That is the frame for Latitude's current product push. The company says its open-source platform captures agent sessions, latency, cost, tool behavior and user-level patterns, then tags unusual behavior and surfaces signals such as satisfaction, frustration, escalation and churn. It also says teams can receive Slack alerts when an agent fails or behaves unexpectedly, then use its MCP server to push fixes from a coding agent workflow and turn real production failures into datasets for regression testing.

The pitch moves Latitude away from generic LLM monitoring and toward what Miguelanez called "conversation intelligence" in the thread. That distinction matters. A normal observability stack is built to inspect systems: traces, spans, requests, errors and performance. Agent products are different because the failure often lives in the gap between what the user expected and what the agent did across a multi-turn exchange. A tool call may succeed, an API may return quickly, and the final user experience may still be wrong.

Latitude's own documentation describes the platform as an open-source, MIT-licensed system for improving production AI agents by capturing real traffic, finding behaviors across that traffic, and turning repeated failures into trackable issues. The docs list the core observability objects as traces, spans and sessions, including LLM calls, tool calls, retrieval steps, HTTP requests, model names, tokens, latency, cost and errors. The newer product language adds the missing commercial layer: not just what did the model do, but what did the customer experience say about the business process wrapped around it.

That is also where the product's incentives show. Agent companies need to convince customers that agents can be operated like software, not supervised like interns. The harder an agent is to monitor, the harder it is to deploy in support, onboarding, sales engineering, implementation or other workflows where the agent may be the first line of customer contact. Latitude is selling a loop: observe live sessions, mine them for failure patterns, group recurring problems, alert the team, and turn those examples into evaluations so the same issue does not quietly return.

Miguelanez has been circling this reliability problem as Latitude has shifted shape. In 2024, Latitude was described by technical founder Gerard Clos as a rebooted open-source product after years of building B2B SaaS behind closed doors. At that point, Clos described Latitude as a free open-source framework for embedded analytics that could create API endpoints on top of a database or warehouse using SQL and embed visualizations in a frontend. Cesar's own maker note from that period said the team had spent two years building software for data teams before rebuilding and open-sourcing Latitude.

By 2025, the product had been recast around agents. In a Product Hunt post, Miguelanez said Latitude had started as a prompt engineering platform and that the agent wave exposed a different problem: people wanted to build agents, but the process was too technical and fragile for many real users. Latitude 2.0's pitch was agent creation and deployment from a single prompt, with observability, A/B tests, synthetic datasets and sub-agent evaluation for technical users.

Tuesday's thread narrows that again. Latitude is no longer only saying it helps teams build agents. It is saying the most valuable operational data appears after the agent meets users. Miguelanez wrote in the replies that "real failures make the best evals," a concise version of the company's operating thesis: evaluation should not be invented in a lab, it should be extracted from the places where production agents actually disappoint users.

The implementation details put Latitude in the same crowded budget line as agent monitoring, LLM observability, evals and product analytics. The company's homepage says Latitude analyzes completed sessions to extract what the conversation was about and flags moments such as escalations, resolutions, abandonments, trust breaks, retries and tool failures. It also says semantic search runs across all traces, with exact text search and metadata filters layered on top. Its GitHub repository describes Latitude as open-source AI agent monitoring, lists a MIT license, and showed about 4,200 stars when checked Tuesday.

Latitude is not giving the market a new metric to worship. It is trying to replace a thin metric stack with a workflow. A customer-support agent that deflects tickets but leaves users angry is not a success just because resolution time improves. A sales assistant that routes a lead to the wrong product may look functional in a trace and still damage conversion. A coding agent that burns through a context window may complete a task while hiding cost and quality problems. Latitude's bet is that the raw conversation can reveal those failures earlier than a customer complaint or churn report.

The pricing also shows the wedge. Latitude's pricing page lists a free Starter tier with 20,000 credits per month, 30-day data retention and unlimited seats, a Pro tier at $99 per month with 100,000 credits and 90-day retention, and an Enterprise tier for higher-volume or self-hosted deployments. That is a developer-led adoption model first, with enterprise security and retention as the expansion path.

The unanswered question is whether conversation intelligence becomes a standalone buying category or a feature inside broader observability and product analytics suites. Latitude's answer is to stay open-source, expose the workflow in the developer surface area, and connect to MCP so fixes can happen where engineering work already happens. That is the product bet inside Miguelanez's thread: for teams running agents in production, watching traces is no longer enough if the real failure is hiding in what the user had to say twice.