Andy Garcia · freelance dev
FRENES
Back to the blog
Careers & AI

HackerRank's New AI Interviewer Lets Candidates Use AI, Then Asks What They Changed in Its Code

October 7, 2026 6 min read

On Monday, October 5, HackerRank made Chakra generally available to its customers. According to TechCrunch, the AI interviewer had spent around six months in beta and conducted more than 500,000 interviews, with Snowflake, Snorkel and Capgemini among the companies that tried it.

Chakra does not read questions from a list. It opens a workspace with a real code repository and an AI assistant, watches the candidate work, then asks why they did what they did. HackerRank's CEO, Vivek Ravisankar, put the reasoning in one line to TechCrunch: "The previous modality of evaluation was evaluating the output. Now, because of AI, anybody can produce an artifact."

I sit on both sides of this table. As a freelance developer, I pass technical screens to win missions. As CTO of CBlindspot, a LegalTech I build on Claude, I decide who joins. So I read past the launch story, into the documentation HackerRank wrote for each side.

What the interview looks like

The hiring team starts with an interview plan generated from a job description: sections, topics, criteria for a strong answer. Every candidate for the role gets the same plan. Chakra can rephrase and follow up, but the bar does not move.

The hands-on part comes in two forms: a coding question solved without AI, or an AI-assisted question in what HackerRank calls an "Agentic Development Environment", where you plan, build and review with an assistant. The demo on the product page shows the second kind: a fictional app called Shipway, a folder of bug tickets, a 30-minute clock, and an instruction to write a PLAN.md before touching any code, covering each problem, the approach, acceptance criteria and risks.

During the build, Chakra stays mostly quiet, and the candidate guide says silence is not penalized. The real interview starts after Submit. The follow-up asks how you scoped the requirements, why you made key decisions, what you reviewed, challenged or changed in the assistant's output, how you validated the result, and how your solution would change if an important constraint changed.

Ravisankar told TechCrunch this single session replaces three rounds: a recruiter screen, a take-home assignment and a follow-up interview with an engineer.

How it scores

HackerRank describes three agents: one turns the job description into the plan, one runs the live interview, and a Reporter reads the full transcript afterwards and scores each expectation on its own. The scale is 3 for met, 2 for partially met, 1 for not met, and 0 for "Not Assessed": "If there's no relevant moment, the expectation isn't scored at all." Scores are then normalized to 0 to 5 for the recruiter. A strong answer in one area does not compensate for a weak one elsewhere.

For the hands-on part, three things are graded separately. The solution, on completeness, root causes, edge cases and regressions (failed test runs and earlier attempts do not count against you). The follow-up discussion, on how clearly you defend the work. And AI fluency: how you frame the task and its constraints, how deliberately you guide the assistant, how carefully you review, test and correct its output. That last score reads both your exchanges with the assistant and your actions in the editor, such as running tests.

The sample report on the product page gives a candidate 4.1 in "Agentic Coding", then explains: "Overall this was passive vibe-coding: the outcome is plausible, but the candidate did not show they understood or could defend what they accepted". The candidate "leaned on 'i trusted the AI'". That sentence is the whole product thesis.

The choice I respect most is "Not Assessed". I run a small judge model for yes/no calls in my own systems, and a score is only worth the evidence behind it.

Cheating, turned around

Allowing AI should make cheating easier. HackerRank says the opposite: suspicious-activity flags were 70 to 80 percent lower than in comparable traditional HackerRank assessments, varying with geography and seniority. Ravisankar's explanation: when an assistant is provided, there is less reason to sneak in another.

It is not an open book. The candidate shares screen and webcam, works full screen on a single monitor, and Chakra pauses if they leave full screen, plug in a second display or leave the camera's view. Screenshot analysis looks for external AI assistants; the product page names apps like Cluely, an external device, or your "friend". The rule moved from "no AI" to "our AI, in plain sight". That is closer to how I work every day.

What it cannot see

The candidate guide says Chakra "does not infer skills or experience that you do not discuss." Fair, but the score then measures explanation under a clock, in a voice conversation. A developer who thinks well but narrates badly pays for it, and possibly so does a non-native speaker, me included. HackerRank says it tests scoring consistency across phrasing, tonality and grammar. I would still check it on my own candidates.

Consistency is not correctness either. A rubric applied identically to everyone applies its mistakes identically too, and the employer writes the rubric. TechCrunch notes that automated hiring tools can inherit bias from their data, models and criteria.

Thirty minutes in a fictional repository will not show how someone behaves after three weeks in a messy codebase, which question they ask a client before writing code, what they decide not to build, or how they disagree with a colleague. In TechCrunch's account, Ravisankar himself leaves to humans the question of whether they actually want to work with a candidate.

Integrity signals are probabilities too. HackerRank's July release added gaze detection, which flags repeated look-aways followed by typing as a medium-severity signal "so hiring teams can review rather than auto-flag". People look away to think. The Chakra report docs say to review the evidence before drawing conclusions.

What I would practise as a candidate, starting Monday

Write the plan before the code, with acceptance criteria you could actually test. When the assistant proposes something, decide out loud, or in the plan, what you keep and what you reject, because "what did you change in its output?" is coming. Run your tests where they can be seen. Answer with your role, the decision, the trade-off and the measured result; the guide says longer answers do not score higher. Say "I don't know" when it is true. Prepare for "what if this constraint changed?". And unplug the second monitor.

What I would keep human as a hiring team

The rubric, first. The July release lets you mark must-have requirements that weigh more in the score, so choosing them is a decision, not a setting.

Calibration, second. HackerRank says it only uses a scoring model once agreement between human experts and the AI is above "a certain level", without giving the level on that page. Check it yourself on Monday: have two engineers score a handful of recorded sessions blind, then compare with Chakra.

Every integrity flag, reviewed by a person before it costs anyone a job. And the data question: the FAQ says candidate data is never used to train models, but that page does not say where webcam and screen recordings are hosted. For a European employer, that is the first question.

The final decision, last. HackerRank's candidate guide says "Chakra does not make hiring decisions", and it should stay that way. New York City already requires bias audits and candidate notice for certain automated hiring tools, and the EU AI Act lists systems that evaluate job candidates as high-risk in its Annex III.

I do not hire at the volume Chakra is built for. But I am borrowing the format: a real repository, the assistant allowed, a plan first, and twenty minutes at the end on what the candidate rejected from the AI. The artifact is cheap now. The explanation is not.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk