How Pendo Uses Claude Managed Agents to Fix Usability Issues Automatically

Pendo has long given product managers the dashboards and analytics they need to understand how users interact with software. But the role of the product manager has been changing fast: AI coding tools have made it possible to ship software faster than ever, which means more code is reaching production without the validation it would have received before. The feedback loop has not kept up with the velocity.

Pendo’s answer to this problem is Novus, an AI-native product that monitors deployed software, detects usability issues in real time, and suggests code-level fixes without slowing teams down. Before Novus, product teams knew something was wrong from drop-off metrics but had no automated path from data to diagnosis to fix. Addressing a single funnel issue required manual investigation across session replays, click data, and code.

Pendo evaluated several approaches before settling on Claude Managed Agents as the foundation for Novus. Earlier attempts at building the agent infrastructure themselves consumed three months of engineering work. Session management had to be built by hand, shared memory required custom plumbing, and all routing logic had to be defined explicitly. When Pendo discovered Claude Managed Agents, three days of integration replaced three months of custom work. Session management, shared memory, and sandboxing were handled out of the box.

Novus works by having users link their GitHub codebase and install a lightweight snippet that captures clicks and session replays. When the agent detects an anomaly—say, a 3% conversion rate on a checkout page receiving a thousand clicks a day—it reviews the relevant code, inspects the funnel data and session replays, and generates a pull request with a proposed fix. The system manages over 150 tools autonomously, routing across them without requiring Pendo to pre-define every decision path. Over 90% of customers who use the MCP integration run it inside Claude Code.

The evaluation results were decisive: when actual product managers reviewed agent outputs line by line, the agent exceeded a 90% success rate—a bar the team considers harder than any synthetic benchmark because the reviewers are the same people who would reject the output in production. Pendo has since extended Claude Code access to its entire engineering team and beyond, with finance and people analytics staff using it for dashboards, board memos, and financial queries.