My testing agent remembers what it learned. Mostly.

说明与用途

一个 AI QA 代理读取 Jira 工单，生成测试步骤，在真实浏览器中执行，比较截图并更新工单状态。新版本存储可用脚本并重用，删除失效脚本，提升执行速度并节省令牌。

问题

Automated testing of Jira tickets in a real browser was slow and costly because each step was regenerated from scratch, even for repeated actions.

结果

The agent now reuses working automation code, discarding broken code, which reduces token usage and speeds up test execution while maintaining accuracy.

工作流

• 1. Orchestrator receives Jira ticket ID, URL, and user email. 2. Orchestrator starts a browser session via BaaS and logs in. 3. Orchestrator fetches the Jira issue and comments. 4. LLM generates an ordered list of test steps in strict JSON. 5. For each step, orchestrator calls sub‑agent step_execution. 6. Sub‑agent takes a screenshot and HTML snapshot, identifies the page. 7. Sub‑agent checks memory for a matching step. 8. If found, reuse stored program; otherwise generate new program. 9. Sub‑agent runs the program in the browser, takes a post‑action screenshot. 10. LLM compares screenshots to expected result and returns verdict. 11. Sub‑agent updates memory: stores successful programs, deletes broken ones. 12. Orchestrator aggregates results, posts a comment to Jira, and if all passed, moves ticket to Done. 13. Browser session is terminated.

证据

An open-source agent that reads a Jira ticket, writes test steps, runs them in a real Chrome window, judges the result from screenshots, and reports back to Jira.

来源

The new version remembers automation code that worked and reuses it when a similar step comes up again, even on a different ticket and phrased differently.

来源

The orchestrator calls the sub-agent as a tool, the same way it would call any function.

来源

If everything passes, it moves the ticket to Done.

来源