Free daily practice for people who build AI agents
The Agentic AI Interview
Real agent runs that went wrong, in the interview that gets you hired as an AI Engineer. The interviewer talks you through the trace, it pauses, you tap the line, say why in your own words, take one follow up, then write the fix and a real model runs it. A score out of 30, and the three things a senior engineer would have said.
About 12 minutes. On a phone you can answer by speaking. Not ready? Learn it first: seven parts, about ten hours, free, or follow the path for your title: AI Engineer, Agent Engineer, AI Evals Engineer, AI Product Manager.
The questions in this set
Each one is a real run from a working agent, with the bug left in. The set grows by one question a week.
| No. | The run | Pattern it tests |
|---|---|---|
| 1 | The double charge. A support agent reads billing tickets and replies to customers. One prompt does everything. | Prompt Chaining |
| 2 | The release note with the wrong number. An agent turns the weekly metrics table into a product update and posts it to the company channel. | Reflection |
| 3 | The tool called with the wrong argument. A shop assistant agent can look up orders with one tool. A customer asks where their order is. | Tool Use |
What agent interviews actually ask
From the published interview guides for AI Engineer, Agent Engineer and Forward Deployed Engineer roles in 2026. The questions are about production failures, not frameworks. Every one of them is a trace in this interview or in the 21 patterns.
- How do you stop an agent that repeats the same tool call?
- Step caps, repeated call detection, stop conditions.
- A tool returns a 500. What does the agent do?
- Retries, fallbacks, and telling the user the truth.
- Untrusted text tells the agent to send data somewhere. Then what?
- Prompt injection, least privilege, what the agent may touch.
- When does one agent beat a team of agents?
- Cost, latency, debuggability, and when an agent is the wrong answer.
- Walk me through a production failure, step by step.
- Trace reading: which tool, what came back, what the model did with it.
- How big was your eval set and where did the cases come from?
- Hand labeled cases from real transcripts and a judge you calibrated.
How it is scored
Ten points a question. Three for tapping the right line on the first try, three for naming the cause and what you would change, one for the follow up, two for a fix that passes a real run on inputs you have not seen, and one senior point for the thing an experienced engineer says first. Nothing is graded by keyword matching alone: the run is real, and the spoken answers are read by a model that never sees the reference fix.