Software quality · Case study
Defects Fixed in minutes can move up.
An agentic quality engineer. It finds the root cause, writes a bounded fix, scores it, stops for a human, and only then writes and runs the tests.
Defect → Fix → Reviewed
Nothing ships until a human says so.
The winning fix, its score and its diff go to a reviewer. Nothing ships without approval. A rejection is a valid outcome, not a failure.
Generic AI code tools make it worse: unconstrained patches, no boundary, no oversight. This one is bounded by architecture, human-gated by design, self-scoring, and regression-complete.
Case study · Defect triage and resolution
Client work
Defects fixed in minutes, inside a boundary the fix can’t cross.
Built on a client codebase
The problem
On a mature app, fixing a bug is mostly archaeology, not repair: locating suspect files, tracing frontend state to backend endpoints, working out what else will break. Hotfixes to shared modules cascade into regressions nobody expected. Under release pressure the tests get skipped, so the bugs come back. Generic AI code tools make it worse: unconstrained patches, no boundary, no oversight.
What we built
An agentic quality engineer that works in five steps. An AST and keyword scan maps the issue to real code symbols, confirmed against evidence. It computes the blast radius before a line is written: files it may edit, callers that must keep working, shared boundaries it can’t touch. It writes a patch blueprint, scores it 0–100 with a deterministic evaluator, and loops up to three times until the score clears the bar. Then it stops for a human. Only after approval does it write and run a four-tier Playwright suite: reproduction, happy path, blast-radius regression and edge.
Bounded by architecture, not hope. The fix stays inside a computed blast radius. Protected modules are off-limits by construction.
Self-scoring, not vibes. A deterministic evaluator decides what’s good enough: a number you can audit, not a hunch.
Human-gated, by design. No code change and no tests until a named person approves the top candidate.
Business outcomes
Client codebase
< 2 min
Root cause, across frontend and backend, against 4.5 hours.
Client codebase
< 15 min
Report to a reviewed, tested fix, against 1.5–2 days.
By design
Contained
Every fix that ships stays inside its boundary.
Why it holds
It writes the tests that stop the bug returning, the step that gets skipped under pressure. Results are from the client codebase this was built on; they reflect that deployment, not a universal guarantee.
Stack
AST and keyword analysis · Blast-radius mapping · Deterministic evaluator · Human approval gate · Playwright