ScyllusWhat this is
Nobody reads the resumes. One opening brings in two thousand by Tuesday, and nobody can read that pile. So we stopped asking anyone to. Scyllus runs AI-assisted coding screens instead: an hour or so of real work on a real codebase with an agent at hand. Some call it vibe coding. There is no leetcode. The screen keeps the whole exchange, prompts, diffs, test results, and it shows what a resume never could: how someone reasons about a problem, how well they direct the AI, and whether they can still fix what it breaks. The hiring team reads a record like the one below and decides. Engineers get read.
From an AI-assisted coding screen · turn 9 of 15 · 41:07 elapsed
| prompt | “Reads come back stale right after a write. My guess: the write path never invalidates. Trace set() in cache.py and show me where the key gets built before you touch anything.” |
|---|---|
| agent | Traced it: set() hashes the namespaced key, get() hashes the raw one. Proposed a one-line patch to set(). |
| check | Rejected the first patch as symptom-only, had it share one key builder between both paths, and required a regression test before accepting the diff. |
| diff | +14 -3 · cache.py, test_cache.py |
| tests | 19 of 19 passing |
A worked example. The record shows how you direct the work, not just whether tests pass.