Pendo · Project Novus
3 months · zero to commercialAn autonomous product engineer, running on the Agent SDK
Novus reads a customer's codebase, instruments their product, watches their users, and ships improvements — continuously, autonomously. The whole loop is Claude: a constellation of subagents coordinated through the Agent SDK, with tool use extending into the customer's repo, their Pendo instance, and their deployed app.
“The most exciting thing to happen to Pendo in years.”
— Pendo Board
After 12 years of customer feedback, Pendo faced a fundamental truth: teams weren't turning analytics data into product improvements. Even with powerful dashboards, most customers needed hand-holding from TAMs, CSMs, and consultants to get value. Project Novus flipped the model — instead of giving teams data and hoping they act, build an AI that understands the product, watches how users behave, and autonomously ships the improvements itself.
Strategic Context
Pendo was absent from AI analyst radars. The board saw a window to transform the company — not just add AI features, but become a recognized AI company. The goal: a new AI-native product line worth tens of millions in revenue.
They needed a team that could go from zero to commercial product in months, not years. No internal ramp-up, no committee-driven roadmap. Just build the thing and ship it.
What We Built
From dashboards to an autonomous product engineer
The old model: give customers analytics dashboards, hope they look at them, then send consultants when they don't. The new model: an AI that reads the codebase, instruments the product, watches how real users behave, identifies what's broken or underperforming, and ships improvements — continuously, without waiting for a human to interpret a chart.
Novus doesn't just collect data. It acts on it. It's the product engineer that never sleeps, never misses a pattern, and always ships.
How Novus works
Understands the customer's product
AI reads the codebase and automatically classifies the business model, detects account structures, and maps the entire data model — no manual configuration needed.
Installs itself
Generates a ready-to-merge pull request that installs Pendo — fully configured for the customer's specific stack and architecture.
Identifies every page and feature
Maps the entire product surface automatically. When tested on Pendo's own codebase, it identified 340+ pages with 95% accuracy — matching what took humans weeks to tag manually.
Instruments the right events
Selects which user actions matter most and instruments them automatically. Customers get meaningful analytics data from day one, not after months of manual tagging.
Continuously improves the product
This is where Novus goes beyond any analytics tool. It watches user behavior, identifies friction and drop-off, and autonomously ships improvements — new guides, UX changes, optimized flows. Not a report. Actual product changes, deployed automatically.
A fundamentally better approach to product analytics
Traditional analytics tools read the DOM — the rendered page a user sees. But the DOM varies by user, subscription level, feature flags, and even browser zoom. It's an unreliable source of truth.
Novus reads the code instead. The codebase is the single source of truth for what a product actually does, and AI can interpret it at a scale no human team could match. This is what makes the 95% accuracy possible — and what makes Pendo's competitive moat defensible.
A new category, not just a new feature
Novus isn't a feature bolted onto Pendo — it's an entirely new product that creates a new category. An autonomous product engineer that any company can plug in and let run. Markets Pendo couldn't reach before — smaller teams, AI-native startups, companies without dedicated product ops — are now accessible.
It also opens new distribution channels. Every app built on platforms like Vercel or Replit could ship with Novus built in from day one — transforming Pendo from a tool you configure into intelligence that's already there, always improving.
The Claude Stack
Orchestrator
Agent SDK · long-running loop
A supervising agent owns the session, delegates to specialized subagents, and checkpoints state so a run can span hours across a real codebase.
Reasoning Tier
Opus · extended thinking
Used for the improvement writer — the step that proposes actual UX and code changes — where depth of reasoning matters more than tokens per second.
Bulk Tier
Sonnet · primary workhorse
Codebase reading, page classification, event selection. Runs the majority of the agent turns with prompt caching keyed on repo context.
Fast Tier
Haiku · routing + triage
Filters incoming events from the deployed product, deciding what’s worth waking the heavier agents for.
Tools
Repo · Pendo API · PR creator
Read/grep over the customer’s source, write against Pendo’s instrumentation API, and emit ready-to-merge pull requests through a Git tool.
Caching
Prompt caching on the codebase
Parsed repo snapshot is cached per session; subagents reuse it across dozens of turns, collapsing cost and latency on the hottest context.
The Agent Loop
Six subagents, one supervising loop, continuous improvement in production
Repo reader
A subagent walks the customer’s codebase using filesystem and search tools, classifies the business model, and builds a compressed map of routes, components, and data flow.
Installer
Takes the repo map, generates the Pendo install configured for the customer’s stack, and opens a pull request through a Git tool. The PR is ready to merge — no humans needed to patch together snippets.
Page mapper
Walks routes and view components to catalog every page and feature. 340+ pages found in Pendo’s own codebase at 95% accuracy — the same tagging that took humans weeks.
Event picker
Looks at the classified product and chooses which user actions matter. Emits instrumentation through the Pendo API so analytics data is live from day one, not after months of manual tagging.
Behavior watcher
Streams user-level events from the deployed product. Haiku triages the firehose in real time; only patterns that look like friction or drop-off get escalated up the stack.
Improvement writer
The hardest step. Given a signal of friction, Opus with extended thinking reasons about the UX, drafts a change (new guide, flow change, copy fix, code patch), and emits either a Pendo guide or a PR through the same Git tool used in step 2.
Why Claude
The loop has to hold together for hours
Novus isn't a single-turn assistant. A real session spans a full codebase, a live product, and dozens of tool calls — and the whole thing has to land in a PR a human would actually merge. That demands a model that tracks state across long horizons without drifting, and tool-use reliability that doesn't silently half-complete.
The Agent SDK is what lets us ship this as a real product rather than a demo: subagents for parallelism, first-class tool use, and prompt caching on the codebase so cost doesn't blow up as the session grows.
Validation & Market Response
Customer Onboarding
Design partners went from connected repo to actionable analytics in a single session. Teams recognized their own pages, features, and user funnels immediately — no training or onboarding required.
Accuracy
When tested against Pendo’s own product — a complex enterprise codebase — the AI matched 95% of page tags that took humans weeks to create manually. 340+ pages identified in under a minute.
Go-to-Market
Phase 1 targeting sub-50 person companies comfortable with cloud access. Same startup segment as PostHog. Sales teams can now set up prospects live during phone calls — from first touch to instrumented product in minutes.
Ecosystem Play
Integration pipeline with Agent Space, Glean, and N8N for PM workflow tools. Plus a future revenue stream through AI-generated app platforms like Vercel and Replit.
The most exciting thing to happen to Pendo in years.
Built on Claude. Designed to compound.
LightCI partners with companies shipping production agents on the Claude Agent SDK.
Talk to LightCI