Stripe's AI Programming Exercise Is a Preview of Every Hiring Loop's Future
Stripe has added a new round to its SWE onsite: the AI Programming Exercise. It runs on HackerRank with a built-in AI assistant. Candidate threads on Reddit are already dissecting it. That's the real signal here — not the round itself, but how unprepared everyone still is to be evaluated this way.
What the round actually looks like
Based on candidate reports, the setup is consistent:
- A HackerRank environment with a lightweight AI chat panel built in.
- A multi-part problem — one recurring pattern is 2 to 3 parts on the same codebase.
- A detailed README the candidate has to read, and can use AI to summarize.
- Roughly 30 minutes of actual coding time — tight enough that hand-coding the whole thing isn't realistic.
- Design and architecture, testing, and optimization are the main evaluation pillars.
Tips: the winning pattern most candidates describe, at Stripe or elsewhere, is having the AI read the README, ask for a summary, and based on that ask for an implementation plan with your input on the approach.
Let AI do the code writing, but spend time reviewing and editing the plan. Sometimes you're expected to add your own tests and edge cases, run them, and be ready to explain what happened. Some candidates reported test cases were already given and they were expected to pass them.
Reviews are mixed, but one candidate who did a similar AI-assisted round elsewhere noted the interviewer explicitly said to skip plan mode since it's slow — just use agent mode.
Be careful — sometimes each part asks for something specific, and giving the AI the entire README leads it to over-build beyond the ask. That's exactly what this round is testing.
Why this round is here to stay
Agentic AI workflows are already normal inside engineering orgs, and hiring is catching up — but the resources to actually prepare for it are thin.
The Reddit thread makes the failure mode obvious without meaning to. Commenters flag that AI output is non-deterministic — same prompt, different result — so you need to really know what you're building and practice prompting.
But hallucination is part of the test. That's the whole point: if two candidates get different AI behavior from the same prompt, the round isn't testing whether you can use AI to solve it, but whether you can detect the issues. That's a stable, evaluable skill.
What good evaluation for this era looks like
The pattern across every comment thread on this round converges on the same three things:
- Spec comprehension under time pressure. Can the candidate extract what's actually required from a README before touching code?
- Verification discipline. Do they write their own tests and edge cases instead of trusting whatever the AI produced, and can they trace a failure back to its cause?
- Legible reasoning. Interviewers repeatedly say they care that the debugging process is readable — narrated hypotheses, not trial and error.
FAQ
What is Stripe's AI Programming Exercise interview round? A relatively new round in Stripe's SWE loop that runs on HackerRank with a built-in AI assistant enabled, designed around using AI as a coding partner rather than testing unassisted coding speed.
What actually happens in the interview? You get a detailed README and a multi-part problem, with an AI chat panel available in the HackerRank environment.
What's the setup — tools, environment, timing? HackerRank with a lightweight built-in AI assistant (candidates compare it to a stripped-down Cursor, so the newest frontier models might not be available). The coding window is short — around 30 minutes — tight enough that most candidates lean on the AI to generate the bulk of it.
What's the step-by-step candidates report as working? Have the AI read the full README, ask it to summarize the requirements back to you, prompt it for a proposed implementation plan and actually review that plan before approving it, let it generate the code, then add your own tests and edge cases, run and debug, and explain everything. Skipping the plan-review step is the most commonly cited mistake — letting the AI run ahead leads to over-building and drifting off spec.
What is Stripe actually testing, if not raw coding ability? Whether you can manage AI output, catch wrong assumptions, write meaningful tests instead of trusting yourself when the plan drifts from the README, and explain your reasoning.
How difficult is it, and how many candidates do well? Public data is thin, but candidate reports are consistent: most people either fully hand the task to the AI and miss edge cases, or try to hand-code too much and run out of time. The multi-part structure means partial completion is common even among reasonably strong candidates — the bottleneck is usually verification and edge-case handling in the later parts, not the first one.
Why does non-deterministic AI output matter for this round? The same prompt can produce different results depending on when you run it. From an assessment-design standpoint, that's arguably the point — testing whether a candidate can catch and correct a bad generation, regardless of what the AI first produced.
Originally published on Medium.