What is a product agent?
A product agent is an AI agent that learns a software product the way a new hire does: it logs in with a normal user account, works through every screen and workflow, and builds its own current picture of what the product looks like and how it behaves. It keeps using the product as it changes, so that picture stays up to date, and it never reads the product's source code.
What a product agent knows
Everything a user could find by clicking around. Which screens exist, what each button and field is labeled, how you get from one place to another, and what happens when you do. Because it learns by doing rather than by reading, it also knows the things nobody wrote down: the settings page that moved last week, the flow that has one extra step for admins, the button that is disabled until you connect an integration.
The picture stays current because the agent keeps using the product. When a release moves something, the next time the agent goes through that workflow it sees the new layout and updates its own map. Nobody has to tell it.
What a product agent can't see
Source code, feature flags, unlaunched screens, internal names. If a logged-in user can't see it, the agent doesn't know it exists, which is a limitation and also the point. An agent that reads the codebase tends to over-share, and it needs access most security teams won't grant to a help tool. A product agent's blast radius is a user account.
Where it's used
Frigade's in-product assistant is a product agent. It answers questions about the product from its own knowledge of how the product works, and takes the action when the question is a task. How Frigade works covers the learning process.
The same agent can keep a help center honest. After every release it does the steps in each article inside the live product, flags the ones that no longer match what's on screen, and drafts the rewrite. Frigade Knowledge is the product that does this.
Common questions
- How is a product agent different from a chatbot trained on docs?
- A docs chatbot knows what someone wrote about the product, which is usually out of date and never covers everything. A product agent knows the product itself, because it has used every screen. It can tell you what a button is called today, walk you to it, and in many cases click it for you.
- Does a product agent need access to my source code?
- No. It only needs a user account, usually with the same permissions a support rep has, on a staging or production environment. Everything it knows comes from what a logged-in user can see and do, which is also why it can't leak feature flags, unlaunched screens, or internal names into anything it writes.
- Knowledge baseA knowledge base is a searchable collection of articles, guides, and answers that explains how a product works, so users and support agents can find answers without asking someone. In SaaS it usually means a public help center plus an internal library that support agents and AI assistants draw on.
- Contextual helpContextual help is assistance that changes based on where the user is and what they are doing in the product, so the answer matches the screen in front of them. It ranges from a tooltip on a single field to an assistant that knows the user's current page and answers accordingly.
- In-app guidanceIn-app guidance is help delivered inside the product interface, where the user is working, instead of in a separate help center, email, or support ticket. It includes tooltips, walkthroughs, checklists, contextual hints, and in-product assistants that answer questions and complete tasks.
- Ticket deflectionTicket deflection is when a customer resolves a question on their own, through self-service help or an automated assistant, instead of opening a support ticket. The deflection rate is the share of would-be tickets resolved this way, and it is the main number used to judge whether self-service is working.