AI-Assisted Coding Rounds

A new round exists, and it is spreading

Between late 2025 and 2026, a genuinely new interview format appeared: a coding round where you are expected to use an AI assistant, live, while an interviewer watches how you work with it. Two major implementations to know:

Meta piloted its AI-enabled coding interview in October 2025 and is expanding it to most backend and ops roles in 2026. The setup: 60 minutes in a custom CoderPad with a three-panel layout — file explorer, code editor, and an AI chat. You work on a multi-file project, not two isolated algorithm puzzles. You can switch among models (GPT-5, GPT-4o mini, Claude Sonnet 4/4.5, Claude Haiku, Gemini 2.5 Pro, Llama 4 Maverick). One deliberate constraint: the AI answers in chat but cannot edit your files — every line that lands in the codebase goes through you. The round replaces one traditional coding round; below E7, it sits alongside a classic no-AI DSA round, so you still need both skills. Meta's own framing: the format "is more representative of the developer environment our future employees will work in, and also makes LLM-based cheating less effective."

Google announced an internal pilot in May 2026 for junior and mid-level US roles on select teams: a "human-led, AI-assisted" code-comprehension round using Gemini specifically — reading, debugging, and optimizing existing code rather than writing from scratch. It explicitly evaluates AI fluency: prompt quality, output validation, and debugging. Context for why: Sundar Pichai said in April 2026 that 75% of new Google code is AI-generated and engineer-approved. Directing and verifying AI output is the job now, so Google interviews for it.

Elsewhere: Canva (since June 2025) expects candidates to use Copilot/Cursor/Claude in technical interviews, and many startups run "vibe coding" prototype rounds. The notable holdout is Stripe, which bans AI in all interview stages including take-homes. Always confirm each company's policy with the recruiter — the same prep does not fit both worlds.

What is actually being scored

Meta grades the round on four axes: problem solving, code quality, verification, and communication. Notice what is not on the list: prompt wizardry. The AI is table stakes; the score comes from your judgment around it.

Interviewing.io analyzed real sessions from Meta's format and found a clean pattern: winners treated the AI "as a teammate" while "gatekeeping code from the model." Concretely, strong candidates:

• Understood the problem before prompting. They read the codebase, formed a plan, and only then delegated pieces.
• Delegated the right subtasks: boilerplate, shell commands, comprehension of unfamiliar code ("what does this module do?"), test scaffolding — while keeping algorithmic and design decisions themselves.
• Critically reviewed every AI output. Read the code, questioned it, tested edge cases before accepting it into the project.
• Narrated constantly. "I'm asking it for a test harness so I can spend my time on the retry logic, which is the risky part" tells the interviewer exactly what they want to hear: deliberate division of labor.

And the documented failure modes — each one fails a real candidate regularly:

1. Pasting the whole problem into the AI and relaying its answer. This demonstrates nothing, and interviewers watch for it specifically.
2. Accepting output unread. One subtly wrong AI-generated function, shipped without review, torpedoes the verification score — which is the axis the round exists to measure.
3. The prompt-tuning spiral. Burning ten minutes rewording prompts to coax a better answer instead of just writing the code yourself. Knowing when the AI is not helping is part of the judgment being scored.

A worked division of labor

Suppose the task is: "This multi-file service has a rate limiter that intermittently rejects valid requests. Find the bug, fix it, and add protection against it recurring." A strong 60 minutes looks like:

• Minutes 0–8, human: read the project structure, trace the request path, form hypotheses. Narrate: "Two candidate causes — a race on the counter, or clock handling in the window reset."
• Minutes 8–15, delegate: "Summarize what token refill logic this class implements, and where state is shared across threads." Use the answer as a map, verify it against the code.
• Minutes 15–35, human: confirm the actual bug yourself (say, a read-modify-write race on the counter), decide the fix. The decision — atomic operation vs lock vs redesign — stays with you, stated out loud with a trade-off.
• Minutes 35–50, delegate + verify: have the AI draft the concurrency test scaffolding, then review it line-by-line, tighten the assertion, and run it to prove the fix.
• Minutes 50–60, human: summarize what was wrong, why the fix is correct, what you verified, and what you would harden next.

The ratio to internalize: the AI wrote maybe 30% of the artifacts, but you made 100% of the decisions and verified 100% of the output.

How to practice

• Recreate the constraint. Practice with a chat-only assistant that cannot touch your files — you type every change. This builds the review-before-accept habit that autocomplete-style tools erode.
• Use multi-file exercises, not LeetCode. Take a small open-source repo or a past side project, give yourself a timed task (add a feature, find a planted bug), and work it with the AI while narrating aloud.
• Practice model triage. Try the same subtask across a couple of models and notice which ones are good at what — a small edge in a timed round.
• Rehearse your workflow answer. Even outside AI rounds, "how do you use AI tools in your daily work?" is now a standard competency question for mid-level candidates. Have a crisp answer: what you delegate, how you verify, and where the tools fail you. That last part — knowing the failure modes — is what makes the answer sound senior rather than credulous.