What is agentic QA?
Agentic QA is software testing done by autonomous AI agents. An agent is given a product and a goal, decides what to test, uses real tools to test it, judges what is a bug, and reports it with proof. Then it remembers what it learned for next time.
The short definition
A QA agent is an AI model (in QA Hub, Claude) running in a loop with tools: a real browser, a shell, an API client, a phone or simulator, or a blockchain test framework. Each session it reads what it knows about the product, plans, acts, observes the result and decides whether something is wrong. Nothing about the test path is scripted in advance.
Agentic QA vs test automation vs manual QA
| Manual QA | Test automation | Agentic QA | |
|---|---|---|---|
| Who decides what to test | A tester | Whoever wrote the script | The agent, from a goal, docs and its own exploration |
| Finds new kinds of bugs | Yes | Only what the script asserts | Yes |
| Survives UI changes | Yes | Often breaks | Adapts |
| Runs around the clock | No | Yes | Yes |
| Explains what it found | Yes | A failing assertion | Steps, expected vs actual, screenshots, recording |
| Gets better over time | Yes | Only when edited | Saves what it learns after every session |
How an agentic QA loop works
- Learn the product. The agent reads docs, user flows, test cases and test accounts, explores the product and writes a model of it.
- Plan. It picks what to test next from an expert playbook for that kind of product (web, mobile, API, smart contracts and so on), recent deploys and open bugs.
- Test with real tools. It clicks, types, calls endpoints or sends transactions, and watches what actually happens.
- Reproduce before reporting. A suspected bug is tried again. Only reproducible issues are filed, with evidence.
- De-duplicate. The same problem seen again is counted on the existing report instead of filed twice.
- Re-verify fixes. When a bug is marked fixed, or a deploy lands, the agent tests it again and confirms or reopens it.
- Remember. Lessons, answers from people and rejected reports become knowledge, so the next session starts smarter.
Where people fit
Agentic QA does not need a person in every step. Teams choose their checkpoints: review every report, review only the uncertain ones, or let reproduced, high-confidence bugs go straight to the board. Agents ask a person only when they are truly blocked, for example by a login wall, and they keep the answer.
How QA Hub does it
QA Hub by Swordfield Technologies gives each product its own Claude QA agents, built from expert templates for websites, AI products, mobile apps, APIs, infrastructure, Web3 dApps, DeFi and smart contracts, games and VR/XR. They test 24/7, on a schedule or on every deploy, file bugs with screenshots and screen recordings, and work with your team from one control room or straight from Claude Code, Cursor and VS Code over MCP. See how it works or try the live demo.
Common questions
What is agentic QA?
Agentic QA is software quality assurance done by autonomous AI agents. Instead of running scripts a person wrote, an agent is given a goal and a product, works out what to test, uses real tools such as a browser, a shell or a device, judges whether the result is a bug, and reports it with evidence. It keeps learning the product between sessions.
How is agentic QA different from test automation?
Test automation replays steps someone scripted in advance, so it only finds what the script checks and breaks when the interface changes. An agentic QA tester decides its own next step, explores flows nobody wrote a test for, adapts when the UI changes and explains what it found. Teams often keep their automated suites and add agents on top.
Does agentic QA replace human testers?
It changes their job. Agents do the repetitive work around the clock: exploring, regression testing, retesting fixes and writing reproducible reports. People decide what matters, set the level of review they want, and answer an agent when it is truly blocked. In QA Hub every human checkpoint is optional.
What can AI QA agents test?
Anything an agent can reach with tools: websites and web apps, AI and LLM products, mobile apps on a device or simulator, APIs and backends, infrastructure, Web3 dApps, DeFi protocols and smart contracts, games and VR/XR apps.
How do you trust a bug an AI agent reports?
By asking for proof. A good agentic QA system reproduces an issue before filing it and attaches steps, expected and actual results, screenshots and a screen recording. It records a confidence level, merges duplicates and lets you choose which reports need a person to approve them.
How do I start with agentic QA?
Pick one product, give an agent the product URL, a test account and whatever docs you have, and let it explore. Start with every report reviewed, then relax the review setting once you trust its reports. QA Hub is invite-only; request access at qahub.swordfieldtechnologies.com/request-access.