After Google I/O 2026 concluded, I sat in front of my computer in a daze for a long time. As a veteran who has been writing code for twenty years, my most intuitive feeling is: The "chat era" of AI is over, and the "working era" has officially begun.
When many people saw the release of Gemini 3.5 Flash, their first reaction might have been: "Oh, another version upgrade, the model is smarter now, right?" But if you carefully read through all the technical documentation, you will find that what Google wants to build this time is not a smarter chatbox at all, but an Agent that can be online 24/7, operate your computer, and run business processes for you.
Today, let's skip the abstract parameters. I will break down in plain English exactly why Gemini 3.5 claims it can serve as your "digital employee."
1. What is an Agent? Why is it More Important Than Chat?
Before discussing Gemini 3.5, let's clarify a core concept: the Agent.
Previously, our logic for using AI was "Q&A-style": you ask a question, and it answers. This is called an assistant. However, the current Gemini 3.5 pursues an "action-oriented" approach: you give it a goal (e.g., "Help me research 10 competitors and write a comparison report"), and it autonomously searches, opens webpages, organizes documents, and sends emails. This is called an Agent.
To achieve this "working like a human" capability, Gemini 3.5 introduces several formidable features:
- Deep Think: Previously, AI was like a "brain teaser," replying instantly without much thought. Now, like a human expert, it drafts several thoughts internally before answering, compares different hypotheses, and selects the optimal solution. Technically, this is called Multi-path Reasoning.
- Computer Use: This is the most "terrifying" aspect. Google's Project Mariner gives AI "hands." It can understand your screen, autonomously click on webpages, drag and drop files, and operate apps. It is no longer just code hiding behind an API, but a "designated driver" that can directly take over your computer.
- Thought Preservation: In previous multi-turn conversations, AI tended to "forget" its prior reasoning logic as the chat progressed. Now, it can remember intermediate reasoning paths, ensuring it doesn't lose track when executing complex tasks.
2. Must-Read for Developers: The "New Rules" of Gemini 3.5
If you are a developer looking to integrate Gemini 3.5 into your own system, the API changes this time are substantial. Google no longer makes you guess its intentions; instead, it hands the control over to you.
Thinking Level
The current API introduces a tiered design, allowing you to choose based on task complexity and budget:
| Level | Scenario | Advantage |
|---|---|---|
| Minimal | Simple chat, quick Q&A | Extremely fast, highly cost-effective |
| Medium (Default) | Most Agent tasks | The balance point between performance and cost |
| High | Hardcore math, complex code refactoring | Most rigorous logic, but slightly slower |
Core Code Example: How to Call Gemini 3.5 to Execute Tasks
Below is a typical Python invocation example. Pay attention to the comments I wrote, as they involve the most core API changes in Gemini 3.5:
import google.generativeai as genai
# Configure your API Key
genai.configure(api_key="YOUR_GEMINI_API_KEY")
# Initialize the model, note that the model name has been updated to gemini-3.5-flash
model = genai.GenerativeModel('gemini-3.5-flash')
# Initiate a chat request
response = model.generate_content(
"Help me analyze this legacy code and provide refactoring suggestions",
generation_config={
# The newly added thinking_level parameter; medium is the optimal choice for most scenarios
"thinking_level": "medium",
# In version 3.5, manually adjusting temperature is no longer recommended; the model will automatically optimize based on the task
}
)
# Print the AI's thought process (if Thought Preservation is enabled)
if response.candidates[0].thought:
print(f"AI's thought path: {response.candidates[0].thought}")
print(f"Final suggestion: {response.text}")
# Note: In 3.5, Function Calling has become stricter
# Every FunctionResponse must carry a unique ID, and the name must match exactly, otherwise the model will throw an error directly
3. How Are Business Processes Executed by Gemini 3.5?
Many business owners ask me: "How much money can this thing actually save me?"
Traditional automation (like RPA) is rigid; you have to script "where to click for step one, where to click for step two." Once the webpage is redesigned, the program breaks. The logic of Gemini 3.5 is semantic-driven automation. It can understand business logic like "account opening" or "reimbursement" and autonomously adapt to different interfaces.
We can look at this Mermaid flowchart to see how a typical "AI employee" works:
mermaid graph TD A[Receive Task: Process customer account opening materials] --> B{Task Breakdown} B --> C[Step 1: Download attachments from email] B --> D[Step 2: Use OCR to recognize ID information] B --> E[Step 3: Log into enterprise CRM system for data entry]
C --> F[Executing...]
D --> F
E --> F
F --> G{Result Validation}
G -- Failure --> H[Deep Think: Why did the data entry fail?]
H --> E[Correct entry strategy and retry]
G -- Success --> I[Send result notification for manual review]
I --> J[Task Closed Loop]
4. Pricing: A 6x Increase, But Why Is Everyone Still Calling It "Worth It"?
There is a point of contention here: the price of Gemini 3.5 Flash is a full 6 times higher than the previous 3.1 Flash Lite. Inputting 1M tokens costs $1.50, and output costs $9.00.
As a veteran, I want to say: Do not just focus on the unit price per API call.
- Efficiency Improvement: The output speed of 3.5 Flash is 4 times faster than its competitors. In Agent scenarios, time is money.
- Higher Success Rate: With previous cheaper models, running a 10-step process might break 3 times, requiring you to repeatedly retry, which also costs money. Version 3.5 is more stable, with a higher probability of succeeding on the first run. Calculated out, the Cost per Task Success might actually be lower.
- Context Window: It supports an ultra-long input of 1M tokens. You can feed an entire thick employee handbook or the source code of a whole project into it, and it will not "lose track."
💡 Summary / Final Thoughts
The release of Gemini 3.5 marks AI's transition from a "toy" to a "tool." If you are still agonizing over how to write a Prompt to make it write poetry, you might truly be falling behind. What you should be considering now is:
- Which repetitive digital labor tasks can be delegated to an Agent? For example, invoice reimbursement, initial resume screening, and code migration.
- Permission Isolation is of utmost importance: Since AI will be operating your computer and enterprise systems, absolutely do not give it "Super Administrator" privileges. It must be run in a controlled sandbox environment.
- Do not blindly trust full automation: At this stage, the most reliable solution is Human-in-the-loop. Let the AI run through the process, and have a human click "Confirm" for the final step.
In summary: The strengthening of models is merely the surface; AI beginning to connect with the real world and execute real tasks is the most hardcore change in this wave. Fellow practitioners, are you ready to welcome your AI colleagues?
