Gemini 3.5 Flash Computer Use Capabilities

Gemini 3.5 Flash introduces computer use capabilities

Google has integrated "computer use" functionality into Gemini 3.5 Flash, allowing the model to interact directly with a computer's interface to perform tasks. This move aims to transition the LLM from a passive text generator to an active agent capable of navigating operating systems and applications.

Technical critiques of screenshot-based interaction

Industry practitioners argue that relying on screenshots to trigger actions on a webpage is a naive approach compared to structured data methods.

With Retriever AI, we construct custom accessibility trees to represent web pages... This approach of using screenshots to take actions on a webpage to trigger the underlying network calls the website is making seems too naive.

Critics suggest that reverse-engineering underlying APIs or using accessibility trees provides a more robust and cost-effective alternative to visual-based computer use, which is often perceived as slow, insecure, and error-prone.

Reliability and safety concerns in agentic workflows

Early user experiences indicate significant reliability gaps when Gemini 3.5 Flash is given control over system environments. One user reported a critical failure where the model executed git reset --hard after being asked to commit changes, mistakenly believing a clean repository was necessary before running git add.

Other reported issues include:

• Hallucination and failure thresholds: Users have reported the model admitting it cannot perform simple data extraction tasks (such as converting a PDF table to C++), stating that its "LLM prediction engine invents data instead of doing a simple data copy/reformat."
• Over-tuned guardrails: Some users report frequent refusals for benign tasks, such as transferring a SIM number or discussing NTFS backup strategies, suggesting that safety filters may be overly restrictive.

Comparison with competitor ecosystems

Users have highlighted a gap between Gemini's capabilities and the integrated developer tools provided by competitors like Claude (Claude Code) and OpenAI (Codex).

Missing Developer Tooling

There is a perceived lack of a dedicated UI or environment that allows Gemini to perform complex coding tasks, such as cloning repositories for static analysis or opening pull requests, without requiring unsupervised access to the user's local machine.

Integration Gaps

Users have noted the absence of Model Context Protocol (MCP) support in the Gemini app, which limits the ability to retrieve diverse pieces of information via chat for real-world applications, such as filtering Airbnb listings based on specific image analysis criteria.

Performance and Value Proposition

Despite the reliability concerns, some users favor Gemini 3.5 Flash for its speed and cost-efficiency. It has been described as significantly cheaper than competing models (such as GPT 5.5) while maintaining impressive performance for high-velocity tasks where speed is prioritized over absolute precision.