Building an agent has never been easier. Anyone with a free afternoon and a vibe coding tool can build one. (We're not judging. We've released weirder things on a Friday.) But making that agent worth using hasn't gotten any easier.
Gartner estimates the cost of that gap: $376.3 billion in AI agent software spend in 2027 alone, with more than 40% of those projects likely canceled — because nobody could prove the agent was working.
That's the exact problem we set out to fix 18 months ago, when we built Agent Analytics. Our goal was to help customers answer one question: is this agent actually helping the people who use it? After more than 47 million agent conversations, we kept hearing the same thing: agents are passing eval checks, but they aren’t actually helping if they don’t know what the user is trying to achieve.. That's the problem Pendo Agent Toolkit was built to solve, and why we're bringing both together as Pendo for Agents today.
Chapter one: Agent Analytics, and what 500 companies taught us
When the first companies started tagging agents inside Pendo, every one of them had an AI product manager asking a question that had nothing to do with uptime: was the agent actually helping the people who used it?
They had session counts and token spend from whichever developer-focused observability tool engineering had installed. None of it told PMs if an agent pilot was ready to move into beta, let alone if the agent could become a revenue-generating product.
So we built Agent Analytics to answer the question developer-first observability tools can't: are users getting value from my agent? And after helping ~500 companies build, deploy, and scale agents, we kept seeing the same four lessons, over and over.
1. Retention decides whether an agent moves to the next stage, not uptime. An agent can have clean logs and pass every eval check, and still not be ready to scale. When Ticketmaster decided whether their agent was ready to leave a limited pilot, the numbers that gave them confidence were retention and rage prompts — not a clean bill of health from engineering.
2. The most important signals usually go unreported. A thumbs-down only comes from someone who takes the time to give feedback, and that's a fraction of the people who are actually frustrated.
The rest show up as in-app behavior (like rage prompting and abandonment). None of it gets caught, because they capture what a user typed, not what happened next.
3. PMs should define success themselves, not just wait for the system to find it. Success doesn't always look like a completed purchase. Sometimes, success means the user got the specific answer they came for, inside the agent conversation. That’s why we built Custom Success Metrics and gave you the ability to create your own use cases and issues: every agent is solving different problems, and you should be able to measure what matters to you.
4. Agent Analytics gets more useful the closer it sits to your code. Some teams are already using the MCP connector for Agent Analytics to build automated loops: Agent Analytics flags the issue, the data flows into the engineering environment, and the fix gets scoped without anyone switching tools. It's the same loop our own team runs on Leo, where it's driven a 61% increase in returning visitors and a 67% drop in issue rate in the last 60 days.
What Agent Analytics showed us next
Even after those learnings, we uncovered another issue: a zombie agent. These agents are technically working — answering prompts and passing every eval check — but have no sense of who they’re actually talking to. The name stuck.
If you open an agent inside your favorite app, you expect it to know you. Nobody wants to explain what they just tried and failed to do, or why they’re turning to the agent for help. So we built something to give agents exactly that: Agent Toolkit, the other half of Pendo for Agents.
Chapter two: introducing Agent Toolkit, the context that turns a chatbot into an advocate
Agent Toolkit gives any agent the product and user context Pendo already captures, delivered to the agent in real time:
- Session context: a real-time record of what someone has done in your product, so the agent already knows that history.
- Frustration triggers: rage clicks, dead clicks, U-turns, and errors, sent straight to the agent so it can step in before someone opens a ticket.
- Workflow nudges: reminds someone exactly where they left off if they stop partway through a specific process.
- Guide surfacing: points to the exact walkthrough, tip, or tutorial that solves what they're stuck on, instead of the agent explaining it from the beginning or sending them to a long help article.
This is the actual fix for a zombie agent. Not a better model. Not a longer system prompt. Context. It works whether your agent is a third-party tool like Fin, Sierra, Zendesk, or Decagon, or something your team built from scratch.
Teachable saw this firsthand. Their support agent, built on Intercom's Fin, had no idea which page a user just left or how many times they'd clicked the same broken button before opening a chat — a textbook zombie agent. Once Pendo started feeding Fin that context in real time, the agent could open with something specific: "Looks like you hit some friction adjusting your settings. What were you trying to change?"
Engagement on proactive messages jumped 200%. 86% of those conversations resolved without escalating to a human, while holding a CX score of 4 out of 5 stars. Best of all, 91% never needed one at all.
Getting started with Pendo for Agents: Less zombie. More advocate.
Today, Agent Analytics and Agent Toolkit live under one roof: Pendo for Agents, built to turn any agent, built or bought, into a customer's best advocate instead of another zombie in the stack. Agent Analytics answers whether an agent is helping. Agent Toolkit makes it better at helping, before the user even enters the chat. One half proves the value, the other half gets it there faster.
If you run support or IT, this is how your agents stop reacting to tickets and free your team for the conversations that actually need a human.
If you own the roadmap, this is how you finally know what your agent should do next.
If you have an in-app assistant or customer support agent or something your own team built, get a demo to see what proactive, context-aware conversations do to your resolution rate.
Already using Pendo? Start using Agent Toolkit for free while in open beta. Simply visit your Pendo platform and enable Agent Toolkit.
