HackerEarth

Scores, reports and integrity signals push "directly into the candidate profile in your ATS," with native integrations to Workday, Greenhouse, Lever, Ashby and more, and custom integrations available via API. That is authenticated write action into a system HackerEarth does not own, carrying its output back to the customer's record.

Coverage stops at one ecosystem class. Every named integration is an applicant tracking system, and nothing documents calendar, video conferencing, messaging, HRIS, e-signature or background check integration. For a product whose function is conducting interviews that is narrower than it first appears, since even scheduling and candidate communication are absent.

Sourcehackerearth.com/ai/onscreenread 2026-09-08

OnScreen conducts the interview rather than scheduling it. A candidate applying at 11pm gets a full interview before Monday morning, through a lifelike AI avatar, in a two way conversation that is dynamic, adaptive and calibrated to the role: it probes depth of knowledge, follows threads and evaluates the quality of thinking behind each answer. Identity verification and proctoring are woven into the same flow, and scores and reports push into the candidate profile in the ATS.

That is one agent running many steps inside a single session on a fixed vendor pipeline (verify, interview, score, report, push), not work passing between distinct participants.

The vendor's own framing points the same way: every interview runs on "a deterministic framework, the same structure for every candidate and no panel-to-panel variation." Invariance is the product promise, the opposite of a branching, configurable flow with many participants.

Nothing documents handoff between the AI Screener, OnScreen and FaceCode, a coordinating layer over them, or a sequence the customer can configure.

Sourcehackerearth.com/blog/introducing-hackerearth-onscreen; hackerearth.com/blog/introducing-hackerearth-onscreen-ai-powered-interviews-around-the-clockread 2026-09-08

The maintained structure is genuine and large: a skills taxonomy spanning 1,000+ skills and 40+ programming languages over a library of 40,000+ problems, with benchmark population data behind the scoring. It persists between runs, scales far past any context window and is queried to calibrate each interview to a role.

It is HackerEarth's own content library, however, not an index over the customer's knowledge. What the customer contributes is rubric and criteria configuration for each role, assembled for each interview. Nothing documents ingestion of the customer's own job descriptions, engineering standards, internal documentation or past scorecards as a queryable grounding layer, and the deterministic framework points the other way: the same structure for every candidate is a guarantee of invariance rather than of grounding in each customer's material.

Sourcehackerearth.com/blog/introducing-hackerearth-onscreen; hackerearth.com/blog/introducing-hackerearth-onscreen-ai-powered-interviews-around-the-clockread 2026-09-08

OnScreen ships one unusual guardrail. Some AI interviewers analyze facial expressions, body language, tone and sentiment, scoring candidates on how confident they seem and how their voice sounds; OnScreen "does none of that, on purpose," on the stated grounds that scoring appearance and vocal style automates the oldest bias in hiring, is scientifically shaky and is a growing legal liability.

That is a shipped constraint on what the agent may consider, and a stronger fairness guarantee than most bias statements because it is a capability the vendor declined to build. Around it, people remain the decision makers on scored evaluations, integrity signals surface in the report for review, and the deterministic framework bounds variation between candidates.

None of that is a review and approve gate over the agent's action. OnScreen conducts and scores a full interview autonomously, and nothing documents an approval queue, a confidence threshold, a hold before a candidate is scored or rejected, or a route for a candidate or reviewer to contest a scored evaluation. Proctoring polices the candidate, not the interviewer.

Sourcehackerearth.com/ai/onscreen and /blog/introducing-hackerearth-onscreenread 2026-09-08

Neither a security attestation nor an access control surface for the buyer is evidenced on the vendor's readable pages. Hiring law compliance with EEOC guidance, NYC Local Law 144 and the EU AI Act is employment regulation, not information security. KYC grade candidate verification and proctoring confirm who the candidate is; they are assessment integrity, not an access control surface for the buying organization, and say nothing about who inside the customer can do what.

The one first party security statement is a 2018 GDPR post saying HackerEarth was "in the process of getting the ISO 27001 certification," which is intent rather than a held certification. HackerEarth serves Google, Amazon, Microsoft and 500+ enterprises, and its product pages render in the browser with no readable footer, navigation or trust surface, so whether a current certification or buyer access control exists is not established.

Sourcehackerearth.com/blog/developers/how-hackerearth-is-preparing-for-gdprread 2026-09-08

The interview produces a structured report rather than a bare number, and behavioral monitoring, detection of AI generated answers and integrity signals surface directly in that report. The deterministic framework means the structure of every interview is knowable in advance, and the vendor's stated purpose is a defensible evaluation. That is a substantive record of the agent's output and the constraints it ran under.

What is documented is the agent's output, a candidate scorecard, not a trace of the agent's own conduct.

The most consequential thing OnScreen does is choose which follow up question to ask next, and nothing published shows which threads the agent chose to follow, which it dropped, or why one candidate was probed on framework design and another on fundamentals.

On an adaptive interviewer that is the auditable surface a buyer would want, and explaining a score is not reconstructing what the agent did. No published scorecard carries evidence at the rubric level, such as quotes, code and timestamps.

Sourcehackerearth.com/ai/onscreenread 2026-09-08

Two facts could look like memory, and neither is. A persistent unified dataset across screening, interviews and development is the application's own data model (candidate records, transcripts, scorecards and assessment history), and deleting it deletes the hiring record itself.

The product is also designed against state carried across sessions: every interview runs on "a deterministic framework, the same structure for every candidate and no panel-to-panel variation." Invariance between runs is the vendor's central fairness claim, and a memory layer carrying learning from one interview into the next would undermine it. Adaptation happens within a session, as OnScreen follows threads and probes depth in response to what a candidate says, which is context inside a single run.

No memory scope, stated lifetime, retention or deletion control over agent memory as distinct from candidate data, or store of taught preferences is documented.

Sourcehackerearth.com/blog/introducing-hackerearth-onscreen; hackerearth.com/blog/introducing-hackerearth-onscreen-ai-powered-interviews-around-the-clockread 2026-09-08

OnScreen is a hosted service that candidates reach in a browser at any hour, and no deployment choice is published. Nothing published names a region, a regional instance, an EU or India hosting option, a customer tenant or a residency selection surface, and no cloud footprint is disclosed either, so there is not even a single named region.

Sourcehackerearth.com/ai/onscreenread 2026-09-08

A product line is published as a browsable top products listing (OnScreen, FaceCode, Technical Assessments, Hackathons, Learning and Development, VibeCode Arena, Job posting and Sourcing), and separately an AI Screener, an AI Interview Agent and a candidate facing AI Practice Agent. Beneath them sits a library of 40,000+ problems with role specific tests and project simulations across 1,000+ skills, a real reusable template bank.

The catalog is thinner than it looks, for three reasons. The named agents overlap: the AI Screener is described as an intelligent, always on interviewing agent that autonomously evaluates candidates against role requirements, near identical to OnScreen's description, and no page distinguishes them.

The AI Practice Agent is interview preparation for candidates, not a unit the buyer adopts. And the problem library is a question bank for assessments rather than a catalog of agents, with no template catalog, agent gallery or clone and configure surface published. The count of distinct agents a buyer can adopt is one, OnScreen, sitting on a genuine content library.

Sourcehackerearth.com/blog/introducing-hackerearth-onscreen, hackerearth.com and /ai/onscreenread 2026-09-08

The trigger is real and it is the product's whole premise. OnScreen exists to close the gap between when a candidate applies and when a person is free, so interviews fire on candidate readiness around the clock rather than on a scheduled slot: a developer applying at 11pm completes a full interview before Monday morning, and one enterprise customer screened more than 2,000 candidates in a single weekend. That is event driven work reaching the agent with no person starting it, with unbounded concurrency behind it.

Channel coverage is narrow and single purpose. The delivery surface is the AI avatar interview session with a live IDE, and the return path is the ATS. No email, SMS, WhatsApp, Slack, Teams, candidate portal or outbound voice channel is documented, so the agent reaches candidates in exactly one place and only once they have arrived there. No trigger framework the customer can configure, webhook or subscription surface is published; the trigger is the one HackerEarth wires into the funnel.

Sourcehackerearth.com/blog/introducing-hackerearth-onscreen and /ai/onscreen; hackerearth.com/blog/introducing-hackerearth-onscreen-ai-powered-interviews-around-the-clockread 2026-09-08

No model picker, admin setting, provider list, routing description or bring your own model option is documented, and HackerEarth names no model provider for OnScreen, the AI Screener or its scoring engine.

One positioning line could mislead. HackerEarth says it evaluates "developers, candidates, and AI agents," using hundreds of millions of evaluation signals to enhance AI model quality for enterprises and AI labs, and VibeCode Arena assesses how a person works with an AI assistant.

A vendor that benchmarks other people's models is not thereby offering the customer a choice of model inside its own product; that is a services line pointed outward. Over the model running the interview agent, the customer chooses nothing and is told nothing. The deterministic framework is a consistency guarantee about interview structure, not a statement about model composition.

Sourcehackerearth.com and /ai/onscreenread 2026-09-08

The OnScreen page names an API: native integrations with major systems, "with custom integrations available via API." That is the vendor naming an API for its own platform.

But the API is asserted without a portal behind it. No developer portal, API reference, endpoint list, authentication guide, SDK or documentation subdomain is published, and the route to the capability reads as a sales conversation about a custom integration rather than a surface a customer builds against unaided. No MCP server is documented.

Sourcehackerearth.com/ai/onscreenread 2026-09-08

HackerEarth evaluates candidates, not its own agent, for the customer. The deterministic, benchmark calibrated evaluation framework with auditable rationale and predictive analytics is HackerEarth's methodology for scoring candidates, not a way for the customer to test the agent. Benchmark calibration against a population is the vendor's own release and tuning practice, not a surface the customer runs or receives output from.

Alignment with NYC Local Law 144 and the EU AI Act is a legal mandate with no audit, auditor, dataset or result named on any first party page.

Nothing a customer could use to test the agent is published: no sandbox, test interview against a known good transcript, versioning of the interview agent, way to exercise a rubric change before it runs on live candidates, or route to an audit of the customer's own configured instance.

Sourcehackerearth.com/blog/introducing-hackerearth-onscreen and /ai/onscreen; hackerearth.com/blog/introducing-hackerearth-onscreen-ai-powered-interviews-around-the-clockread 2026-09-08

HackerEarth controls a computer, but in the wrong direction for this capability.

Its Smart Browser is a lockdown runtime on the candidate's own machine that blocks application and tab switching, copy paste from external sources, second monitors and extended displays over HDMI, screenshots and screen recording, virtual machines and remote desktop sessions, and browser extensions including AI assistants, surfacing violation attempts to reviewers for audit after the assessment.

That is a restraint imposed on the candidate's machine to preserve assessment integrity, not the agent operating a browser or desktop to do work for the customer.

The live IDE and coding sandbox are where the candidate writes code, and sandboxes, terminals and code execution are not computer use. Neither a lockdown runtime nor a proctoring feed drives someone else's interface.

Sourcehackerearth.com/blog/remote-proctoring-vs-smart-browser and /ai/onscreen; hackerearth.com/blog/remote-proctoring-vs-smart-browser-what-each-catches-what-each-misses-and-how-to-chooseread 2026-09-08