Sierra's debugging round can give you an AI agent that won't find the bug

Photo: Kyle Wong on Unsplash

In Sierra's newer debugging rounds, the repo tells the AI coding agent not to find the bugs for you. We found debugging in 11 of 48 interview loops, and an AI coding agent in 5.

In 5 loops you work with an AI coding agent; only Sierra's is told not to find the bugs for you.
11 of the 48 interview loops we map include debugging; only 3 (Sierra, Scale AI and Notion) make it a round of its own.
5 of those 11 let AI into at least some debugging sessions: Nvidia, Sierra, Stripe, Amazon and Uber.
The round is graded on your diagnosis: reproduce, find the root cause, fix, and prove it works.

11 of 48 loops include debugging; 3 make it a round of its own

Loops by how they include debugging (n=48). “One option” means debugging is one of several tasks the round can be (debug, integrate or build at Stripe; a short task in some screens).

It isn’t only an AI company habit: 5 of the 11 are AI companies (of 34 AI company loops we map) and 6 are big tech or scaled startups.

5 of the 11 let AI into at least some debugging sessions

Loops that include debugging, by what their reports say about AI in it (n=11).

Sierra’s AI agent is told not to find the bug for you

Sierra’s onsite has a 60-minute debugging round. A small repo that implements one feature end to end (a Python retention agent in a 2026 report, a TypeScript and React app in earlier ones) with several customer-reported bugs: reproduce each, find the root cause, fix it and prove the flow works end to end. Earlier reports: no AI coding tools. Newer loops can include an AI coding agent whose repo instructions (AGENTS.md) tell it not to find the bugs for you: it acts only once you point it at the direction and the code.

In the version candidates report for 2026, the repo is a retention agent for a subscription business: it offers discounts from two pots by member tier, and three customer complaints come in. The bugs are in which pot a check reads and which pot an offer spends. The agent can run the customers and make the edit, but finding the line is your job. We wrote it up as a practice question with tests for each complaint: Retention agent: fix three bugs in a discount offer engine.

In 4 more loops the agent is there to do the work

Anthropic, OpenAI, Applied Intuition and Physical Intelligence have rounds where you work with an AI coding agent, and their reports say the agent is expected: there, using it well is the work. At Anthropic: some 2026 loops add a separate round where you write and review code with Claude Code in a hosted editor; there AI use is expected.

What this means for your prep

Ask your recruiter whether a debugging round is in your loop and whether AI is allowed in it. In 5 of the 11 loops with debugging, AI is allowed in some sessions but not all.
Practice on code you didn’t write. Read the whole thing once, then reproduce each complaint with one concrete input before you change anything.
Keep a ledger of state. Write the values down after each step; most planted bugs read the wrong field or spend the wrong one.
Use the agent for the legwork, not the diagnosis. At Sierra, it acts only once you point it at the direction and the code, so the narrowing down has to be yours.
Prove the fix end to end with a test per complaint and a full run per case, and check the fix didn’t change another path.
Methodology

The loops of 48 tech companies, round by round, from candidate reports and public guides, checked in November 2026. A loop has a debugging round of its own when a round is named for debugging alone; it includes debugging as one option when debugging is one of several tasks a round can be, or a round’s notes mention a debugging task. AI use comes from each round’s stated policy and notes; AI coding agent rounds are those whose policy or notes mention working with one.

Limits: loops change by team, level and month, and newer formats (AI in the room, agents in the repo) can be pilots that not every candidate gets.

How we made this: Land the Lab sorted each round’s description with keyword rules. The charts are generated from that data; the text was drafted with AI and checked against it.