Gemini 3.5 Flash: Build Site QA Agents for Pennies in 2026

I can now hand a Gemini agent a URL and get back a full UX audit with screenshots in about 90 seconds, for less than a dollar. This isn’t a demo; it’s a billable service I run before every client handoff.

Google just baked its ‘Computer Use’ tool directly into Gemini 3.5 Flash—its fastest, cheapest model. This means building agents that can see your screen, control a browser, and run QA checklists is now practical for any agency.

---

What is Gemini Computer Use?

Gemini Computer Use is a native tool in the Gemini API that allows an AI agent to see a user’s screen, understand the user interface, and control the mouse and keyboard. Integrated into Gemini 3.5 Flash, it enables the model to perform tasks like navigating websites, clicking buttons, and extracting data directly from a browser without needing separate tools or model handoffs.

---

The QA Agent Config I Use — Copy It Now

This is the exact agent configuration I use to run a pre-launch quality assurance audit on a client’s landing page. It combines a system prompt that defines the agent’s role with a user prompt that specifies the URL and the audit checklist. The Python code shows how to call the Gemini 3.5 Flash model with the Computer Use tool enabled.

# 1. THE SYSTEM PROMPT (Defines the Agent's Persona)
"""
You are a meticulous Quality Assurance (QA) agent specializing in website user experience (UX) and functionality audits. Your task is to open a provided URL, visually inspect the page, and report on specific criteria. You must navigate the page as a human would, checking for visual consistency, clarity, and errors. Your output must be a structured JSON object containing a list of issues, each with a type, severity, description, and suggested fix.
"""

# 2. THE USER PROMPT (The Specific Task)
"""
Audit the landing page at the following URL: https://example-client-site.com

Use the Computer Use tool to open the page and perform the following checks:
1.  **Broken Links:** Click every navigation link and call-to-action (CTA) button. Report any that lead to a 404 page or do not work.
2.  **CTA Clarity:** Read the text on all primary buttons. Report if any are vague (e.g., 'Click Here', 'Submit').
3.  **Visual Regressions:** Scan the page for any obvious visual bugs like overlapping text, broken images, or elements that are misaligned.
4.  **Content Gaps:** Identify the main value proposition. Report if the headline and sub-headline clearly state what the product/service is and for whom.

Return your findings as a single JSON object with a key 'issues' containing an array of found problems.
"""

# 3. THE PYTHON SDK CALL (How to Run It)
import google.generativeai as genai

# Configure with your API key
genai.configure(api_key="YOUR_API_KEY")

# Initialize the model with the computer_use_tool
model = genai.GenerativeModel(
    'gemini-3.5-flash',
    tools=[genai.tool.computer_use_tool.ComputerUseTool()],
)

# Start a chat session with the system prompt
chat = model.start_chat()
chat.send_message("YOUR_SYSTEM_PROMPT_HERE")

# Send the user prompt to execute the audit
response = chat.send_message("YOUR_USER_PROMPT_HERE")

print(response.text)

When I ran this on a client’s staging site for the first time, it caught two broken checkout links that the human QA team had missed. That’s the power of exhaustive, automated checks.

+----------------------+    +-----------------+    +-----------------+
|   My Prompt & URL    | -> | Gemini 3.5 Flash| -> |   Live Website  |
+----------------------+    | + Computer Use  |    |  (Browser View) |
                          +-----------------+    +-----------------+
                                  |
                                  V
                          +-----------------+
                          |  Structured QA  |
                          |   Report (JSON) |
                          +-----------------+

---

Here’s Exactly How I Build a Site Auditor Agent

I build a site auditor agent by defining its purpose in a system prompt, providing the target URL and checklist in a user prompt, and then executing the task using the Gemini API with the Computer Use tool enabled. The entire process, from getting my API key to running the first report, takes less than 15 minutes.

1. Get a Gemini API Key. I go to Google’s AI Studio, create a new project, and generate an API key. This is a one-time step that gives my script permission to talk to the Gemini models.
2. Define the Agent’s Persona. I write a clear system prompt, like the one in the asset block above. This tells the agent its job is to be a QA tester, what kind of issues to look for, and how to format its report. This step is crucial for getting consistent, high-quality output.
3. Create the Specific Task. For each audit, I write a user prompt containing the exact URL and the checklist of items to verify. This makes the agent reusable; I just swap out the URL and checklist for each new client or project.
4. Execute the Python Script. I plug the system prompt, user prompt, and my API key into a simple Python script and run it from my terminal. The agent takes over my screen, opens a browser, performs the audit, and prints the JSON report right in the terminal.

---

What This Changes for Designer-Run Agency Work

This changes agency work by turning manual, time-consuming quality assurance into a cheap, automated, and billable service. It allows a single designer to perform the work of a junior QA tester in minutes, reducing project overhead, speeding up delivery, and creating a new recurring revenue stream for site maintenance retainers. This is a key part of my Gemini playbook for agencies.

Dimension	Old way	New way
QA Process	A junior dev or designer manually clicks through every page.	An automated agent runs a pre-defined checklist on a URL.
Cost	$25-$50/hr for a junior team member’s time.	~$0.30 per audit, based on Gemini 3.5 Flash token pricing.
Speed	2-4 hours for a comprehensive check of a 10-page site.	Under 10 minutes for the same 10-page site.
Deliverable	A Google Doc or PDF with manually captured screenshots.	A structured JSON report, ready to be fed into a project management tool.

The biggest shift is that I can now guarantee a baseline level of quality on every single project that goes out the door, without eating into my profit margins. It becomes a standard, non-negotiable step in my delivery process.

---

My $0.02 — How I’d Roll This Out

I’d roll this out by starting with a single, high-value internal task to prove its reliability before offering it as a client service. The goal is to build trust in the automation, create a standardized report format, and then productize the audit into a line item on new proposals and maintenance plans.

Day 1 — Audit my own site. I’d start by running the QA agent on my own agency website, mikekwal.com. This is the safest place to experiment. I’d refine the prompts until the agent consistently finds real, actionable issues and formats the report exactly how I want it. The goal is to get one perfect, end-to-end run.

Day 2 — Standardize the deliverable. I’d take the raw JSON output from the agent and build a simple script to convert it into a clean, client-facing PDF report. This report would include the issue, severity, a screenshot, and the suggested fix. This turns the technical output into a professional asset I can actually sell.

Day 3 — Productize the service. With a reliable process and a professional deliverable, I’d add ‘Automated Pre-Launch QA Audit’ as a standard line item on every new website proposal. I’d also offer a ‘Monthly Health Check’ as part of a maintenance retainer, using the agent to continuously monitor client sites for new issues.

---

FAQ

How much does it cost to run a Gemini 3.5 Flash site audit?
An average landing page audit costs between $0.20 and $1.00. Pricing is based on the number of tokens used, which depends on the complexity of the page and the length of the audit. At roughly $1.50 per million input tokens, it is significantly cheaper than human testing or competing models.

How accurate is Gemini’s Computer Use compared to a human tester?
The OSWorld-Verified benchmark places its accuracy at 78.4%, which is within 0.3 points of GPT-5.5’s score. It excels at finding objective issues like broken links or missing images. It is less effective at subjective feedback on design taste, which is where human expertise remains essential.

Is it safe to let an AI agent control a browser on my machine?
Yes, when managed correctly. The agent only runs when you execute the script and only has the permissions of your user account. For sensitive tasks, I run it in a sandboxed environment or a dedicated virtual machine. Google has also built-in safeguards to stop tasks if prompt injection is detected.

Can this agent test mobile responsiveness?
Yes. You can instruct the agent to resize the browser window to specific mobile viewport dimensions (e.g., 390×844 for an iPhone 14) and report on any layout issues that appear. This allows for basic responsive testing as part of the standard audit checklist.

Does this replace tools like BrowserStack or manual QA teams?
No, it complements them. It replaces the repetitive, checklist-based tasks that are a poor use of a skilled human’s time. I use it for the first pass, which frees up human QA testers to focus on more complex user journeys, accessibility testing, and providing subjective design feedback.

What kind of issues can the agent find?
The agent is best at finding functional and content-related issues. This includes broken links, non-working buttons, 404 errors, placeholder text left in the copy, confusing CTA labels, missing image alt text, and major visual bugs like overlapping elements or broken layouts at specific screen sizes.

Is Gemini 3.5 Flash with Computer Use available now?
Yes, it entered public preview on June 24, 2026. It is available through the Gemini API and the Gemini Enterprise Agent Platform without a waitlist. You can get an API key and start building with it immediately.

Can I use this for more than just QA?
Absolutely. The same Computer Use tool can be used for competitive research (extracting pricing from competitor sites), data entry (filling out forms), or content scraping (gathering testimonials). The QA audit is just one of many potential applications for an agency.

---

Want help applying this?

Four ways to go deeper:

• Build with Builders. Join the Talk-to-Build community to learn to build AI-native websites, cinematic AI video, and agent-driven workflows you can sell.
• 1-on-1 working session. Book a screen-share with me — bring a real problem, leave with a working piece of it.
• Done-for-you. MK-Way builds AEO-ready websites, apps, and AI agent workflows.
• Quick question. DM me on Instagram or LinkedIn. I read every message.

---

Part of the AI Pulse series. If you commented “GEMINI” on one of my videos — this is the breakdown.

Last updated: 2026-07-09.