Browser Workflow QA Agent

AI agents are increasingly used for software quality checks. Most produce unstructured chat output: "the page looks fine", "no obvious errors". That is not QA. It is a guess with no audit trail, no reproducibility, and no way to track regressions over time.

The real problem is that agents default to summarizing what they see rather than capturing evidence of it. A QA workflow that lives only in a chat window is worthless the moment the window closes.

A structured QA agent triggered by a natural-language Telegram message: "Jarvis, QA this URL: [url] Goal: [objective]". The agent executes a deterministic inspection workflow and saves a full evidence bundle. Not just a chat reply.

The workflow terminates with a structured Markdown report saved to durable storage and a concise summary sent back via Telegram. Findings persist whether or not the conversation continues.

Evidence-first, not chat-first. The workflow is designed around what gets saved, not what gets said. Every run produces a Markdown report, a screenshots bundle, and a metadata file. All independent of whether the conversation continues. QA that exists only as chat history is not QA.

Small, testable reporting substrate. Rather than embedding report logic inside agent prompts, I extracted it into a reusable Python helper (qa_report_helper.py) with deterministic slug generation, verdict classification, and structured Markdown output. TDD discipline: tests cover filename safety, verdict logic, required report sections, and output naming conventions.

Natural-language trigger, structured execution. The Telegram trigger is intentionally casual. One line, conversational. The agent's job is to translate that into a deterministic inspection checklist. The interface is human. The process is not.

The boundary between "agent observes" and "agent concludes" required deliberate design. Without explicit constraints, agents tend to collapse multi-step inspection into a single verdict, skipping the evidence capture that makes the verdict trustworthy. The workflow enforces screenshot capture and console logging as required steps, not optional ones.

Structuring findings by severity (high / medium / low / info) rather than as a flat list required judgment: what makes something actionable vs. informational? The verdict schema needed to match what a real QA engineer would care about, not just what the agent happened to notice.

This project demonstrates practical agentic automation applied to a real engineering need: making quality assurance auditable, reproducible, and persistent. It's not a demo that shows what an agent can say. It is a workflow that shows what an agent can prove.

Natural-language triggers, deterministic execution workflows, evidence capture, structured verdicts, and testable helper substrates generalize to any domain where agents need to produce trustworthy outputs rather than impressive-sounding ones.