A toolkit of independent AI agents for QA — each one solves a single real testing problem, built only where nothing good already exists.
Flaky tests, requirement ambiguity, coverage gaps, framework migrations, duplicate test cases, bugs that slip through a green build — each of these is a distinct, specific problem, not a feature of one grand platform.
Most "AI for QA" products pick one of these, wrap it in a dashboard, and call it a suite. AgenticQA takes the opposite approach: treat every problem as its own small, finishable build, and only build what doesn't already exist.
Before writing a line of code for a new agent, three questions get answered in order:
Every agent runs as its own independently deployable service — one can be demoed, shipped, or thrown away without touching the others. One shell app browses and launches all of them; it does not pass data between them.
Only three agents share a genuine dependency: a searchable corpus of existing test cases. That shared piece is built once, as its own service, and called by the agents that actually need it — everything else in the catalog stays standalone.
Pulls a Jira ticket, flags vague wording and untestable clauses, scores confidence, and produces clarified acceptance criteria — shifting ambiguity detection left, before a test is ever written.
A generator + reviewer agent pair. Turns a sanitized requirement into a confidence-scored test suite, explicitly addressing every gap the smell detector flagged — no silent drops, bounded revision cycles.
Given a customer-reported bug: checks reproducibility, searches the test repo for coverage, checks whether a covering test actually ran and passed in that release, and classifies exactly why the bug leaked.
The next agent formalizes a real QA workflow: when a customer reports a bug, the question isn't just "why did it happen" — it's "why didn't our tests catch it." The answer is a specific, traceable classification, not a guess.
Twenty-plus real QA problems, one small finishable build at a time — never one framework trying to solve all of them at once.
Every agent is checked against what already ships natively or open-source first. Building is earned, not assumed.
One launcher app for discovery and demos. Agents stay independently deployable; only real data dependencies get wired up.
Each agent encodes an actual QA practice — defect leakage analysis, requirement ambiguity review — not a generic "AI does X" pitch.