Teresa Torres wrote the book on continuous discovery, and I am a fan. Interview a user every week, build an opportunity solution tree, test assumptions. The playbook is solid and most teams under-use it. It has one hidden assumption: your customer is a human who can be interviewed. In 2026 a growing share of who actually interacts with your product is not human. It is an agent that reads your API, uses your tools, and hits your endpoints. It does not have opinions, get frustrated, or answer interview questions. If agents are heavily using your product, you have a discovery problem the canonical playbook cannot solve.
Know how agents use your product
Four patterns, most products have a mix. The tool-user agent calls your API as one step in a multi-step workflow, like a coding agent looking something up in your docs. The end-to-end replacement agent is the customer, with the human at a higher level of abstraction, like an accounts payable agent paying invoices in your billing platform without a human logging in. The intermediary agent sits between human and product, translating, like a meeting assistant. The hybrid agent-human session has both collaborating inside a session, like an analyst working with a research agent. Each needs different discovery methods, and none is well served by interview a user every week.
Six methods that work
Agent telemetry as primary signal. Instrument what agents do at high fidelity: which endpoints they hit and in what order, which calls retry or fail and what the agent does next, which sessions get abandoned mid-flight. That telemetry is your interview transcript when the customer cannot talk.
Failure-mode interviews with operators. You cannot interview the agent, so interview the human who built it. Ask what agent they built, where it works, where it fails, the ugliest workaround they built, and the one change that would most help. Thirty minutes each, five operators per quarter. The signal density beats a traditional user interview, because the operator already did the work of turning fuzzy needs into a specific configuration.
Capability audits. Monthly, describe a real workflow to Claude or GPT and let it try to execute against your actual API or UI in a sandbox. Log every failure, ambiguity, and retry, then classify each as a product, documentation, or model limit. Two hours of audit reveals more than a week of interviews.
Pattern mining, a benchmark suite, and outcome interviews. Mine GitHub issues, Discord, and forums weekly for what people build and hack around. Stand up a benchmark of 20 to 50 real agent tasks and score it quarterly; a declining score means your product is drifting away from agent-usability. And interview the end-user humans whose work the agent automates, about outcomes rather than features.
The deployer is the new customer
The human who deployed the agent makes the buying decision, configures the tools and constraints, and lives with the consequences. Interview them, but do it to understand, not to optimize. Discovery first, decision second. This week, tag every API call by agent versus human and get five agent-usage signals on a dashboard. Next week, book five failure-mode interviews. The first one usually resets your roadmap.