Real work reveals
real engineers.
On hunr, candidates build real work with any agent — then a live defense proves they understood it.
Published pricing. You only pay the placement fee if you hire through hunr.Any agent allowedNo surveillanceVerified understanding
Evidence-linked · anchored to a reference expert
The signals you hire on stopped working.
Résumés never shipped. The take-home did — until any candidate could hand it to an agent.
Résumés don't predict the job.
Pedigree and LeetCode speed say little about who can navigate a real codebase and own what they ship.
AI broke the take-home.
Candidates already use agents every day. A static assignment now measures the model, not the engineer.
Proctoring is an arms race.
Lockdown browsers and AI-detection add friction, punish honest candidates, and still miss who really understands.
A JD in. A complete challenge out.
Seed it with your business, your stack, and the competencies that matter. The agent builds a runnable repo, hidden tests, reference and naive solutions, and a weighted rubric — you review and set it live.
- app/ · src/
- tests/hidden/ · tests/visible/
- .hunr/rubric.yaml
- solutions/ — reference + naive
- BRIEF.md
- Role- and stack-specific, generated from your context
- Hidden tests are authoritative — you approve before it goes live
- Your competency weights drive how candidates are ranked
Every submission, run for real.
Candidates push to their own GitHub repo. We run it in an isolated sandbox against the hidden gateway suite — pass/fail, reproducible, and hard to fake.
- test_returns_booking_for_windowvisible
- test_rejects_overlap_409visible
- test_partial_overlap_returns_409hidden
- test_concurrent_bookings_racehidden
Hidden tests are authoritative — reproducible, and impossible to shadow.
- Hidden tests are authoritative and can't be shadowed
- Visible Actions feedback while they work; authoritative hidden run on submit
- Screen many candidates without a human grader per run
Then prove they understood it.
A live, unaided reasoning round asks why — trade-offs, failure modes, the road not taken. An ownership multiplier discounts the score when the answers don't hold up. The gap between a passing artifact and real understanding is the signal.
- Question types: rationale, trade-off, failure mode, novel extension
- Ownership multiplier scales the technical score; below the floor, it fails
- One question per turn, time-boxed, scored live
Hire on the signal, not a gut feel.
Passing submissions get a per-competency analysis — every score evidence-linked to the candidate's own code and anchored to a reference expert, so a number actually means something. You get a ranked, defensible shortlist.
See pricingEvery score is evidence-linked to the code they wrote · the marker is a reference expert.
- Your competency weights drive the ranking
- Hard gates eliminate; graded dimensions rank
- Reports you can defend to your team, line by line
A complete challenge, not a prompt.
Plenty of tools spin an assessment from a JD. Almost none make the pieces fit together. The bundle is the hard part.
A complete, runnable repo
A realistic codebase in the candidate's actual stack — not a blank editor or a puzzle prompt. This is the seed everything else hangs off.
Deterministic hidden tests
Authoritative pass/fail that can't be shadowed or gamed.
A matching weighted rubric
The same competencies that grade the work also rank the shortlist.
Built-in reasoning questions
Understanding-verification baked into the artifact, not bolted on — one coherent, traceable bundle.
Every role, every stack.
Each challenge is generated in the candidate's real discipline and stack — backend to mobile, in the languages we grade for real.
Don't see your stack? We generate to it.
Stop guessing. Start shortlisting.
Book a demo and we'll generate a challenge for one of your open roles — on us.