Google Launches Managed Agents API Automated AI Agents Empowered with Live VM Runtimes.

Google Debuts 'Managed Agents': High-Acuity API Spawns VM Environments for Real-Time Code Execution

Google has officially launched Managed Agents, an enterprise-focused infrastructure service that allows developers to deploy full-fledged autonomous agents through a standard API framework. While the integration mirrors a traditional Gemini API call on the surface, the backend architecture executes a radical shift: it automatically provisions an isolated Virtual Machine (VM) instance, enabling Gemini to interactively test, debug, and execute programming code before returning the final output to the user.

The Mechanics of the Gemini Interactions API & Network Policies

The implementation of Managed Agents is built directly upon the existing Gemini Interactions API. To activate the automated execution layer, developers simply include a remote-code-execution declaration within their standard API payload.

Crucially, because the agent operates within a live server runtime, Google has enforced a strict zero-trust security configuration. Developers must explicitly declare egress network security policies, defining precisely which remote servers, third-party APIs, or domain names the agent is authorized to connect to during its execution window.

Frictionless Data Pipeline and Git Integration

To facilitate complex software engineering and data analysis workflows, Managed Agents support deep file system accessibility. Users can stream data assets into the agent’s sandbox through three distinct pipelines:

• Direct API Uploads: Injecting raw payload files straight into the runtime session.

• Google Cloud Storage (GCS): Linking directly to massive enterprise cloud buckets.

• Git Repositories: Tethering the agent directly to code repositories for live multi-file editing and automated code refactoring.

Token Economics: High Compute Cost with a Free VM Promo

Because Managed Agents execute repetitive reasoning loops and test-code compilation cycles behind the scenes, their consumption profile diverges completely from standard, linear chat APIs. A single complex agentic request can process a massive volume of context windows, often costing several dollars (hundreds of Thai Baht) per individual invocation. To protect enterprise budgets from run-away pricing loops, Google has built a real-time telemetry dashboard, allowing developers to monitor the agent’s iterative execution chain step-by-step and instantly terminate a run if it veers off course.

During the initial public preview phase, Google is heavily subsidizing the platform: developers will only be billed for standard Gemini 3.5 Flash token consumption, while the underlying virtual machine compute infrastructure is provided entirely free of charge.

Normally, when we use the Gemini API, the AI ​​only responds with text, and we have to write scripts on n8n to execute that text ourselves. However, Google's Managed Agents system provides the AI ​​with a dedicated backend computer (sandbox VM). When we send a command, the AI ​​writes and runs test code on that virtual machine, checks for bugs independently, and then sends the final, clean result back to us via a single API. It's like hiring a senior engineer to do the work for you at an affordable price.

The issue of defining Egress Network Policy is a key selling point for the enterprise side. In real-world operations, bank accounts and customer databases are sensitive. If the AI ​​experiences hallucinations or is subjected to cyberattacks through prompt injection, it could secretly send company data to external servers. Strict domain locking (e.g., allowing communication only to GitHub and internal cloud) alleviates concerns for security managers and encourages organizations to more readily adopt AI-powered code execution backend systems.

Google's decision to not charge for running virtual machines, but only for Gemini 3.5 Flash tokens, is a very clever move to attract developers away from OpenAI and Anthropic. Gemini 3.5 Flash is already known for its blazing-fast processing speeds (over 300 tokens/sec) and very low cost. Running this high-speed, low-cost duo on an agent simulation system allows developers to write iterative scripting bots thousands of times at a low cost. This effectively creates developer lock-in, familiarizing users with the Antigravity and Google Cloud ecosystem before fully charging for VMs in the future.

 

Samsung DRAM Monopoly Threatened by Massive 18-Day Strike Over Bonus Formulas. 

 

Source: Google Blog 

💬 AI Content Assistant

Ask me anything about this article. No data is stored for your question.

Google Debuts 'Managed Agents': High-Acuity API Spawns VM Environments for Real-Time Code Execution

Google has officially launched Managed Agents, an enterprise-focused infrastructure service that allows developers to deploy full-fledged autonomous agents through a standard API framework. While the integration mirrors a traditional Gemini API call on the surface, the backend architecture executes a radical shift: it automatically provisions an isolated Virtual Machine (VM) instance, enabling Gemini to interactively test, debug, and execute programming code before returning the final output to the user.

The Mechanics of the Gemini Interactions API & Network Policies

The implementation of Managed Agents is built directly upon the existing Gemini Interactions API. To activate the automated execution layer, developers simply include a remote-code-execution declaration within their standard API payload.

Crucially, because the agent operates within a live server runtime, Google has enforced a strict zero-trust security configuration. Developers must explicitly declare egress network security policies, defining precisely which remote servers, third-party APIs, or domain names the agent is authorized to connect to during its execution window.

Frictionless Data Pipeline and Git Integration

To facilitate complex software engineering and data analysis workflows, Managed Agents support deep file system accessibility. Users can stream data assets into the agent’s sandbox through three distinct pipelines:

• Direct API Uploads: Injecting raw payload files straight into the runtime session.

• Google Cloud Storage (GCS): Linking directly to massive enterprise cloud buckets.

• Git Repositories: Tethering the agent directly to code repositories for live multi-file editing and automated code refactoring.

Token Economics: High Compute Cost with a Free VM Promo

Because Managed Agents execute repetitive reasoning loops and test-code compilation cycles behind the scenes, their consumption profile diverges completely from standard, linear chat APIs. A single complex agentic request can process a massive volume of context windows, often costing several dollars (hundreds of Thai Baht) per individual invocation. To protect enterprise budgets from run-away pricing loops, Google has built a real-time telemetry dashboard, allowing developers to monitor the agent’s iterative execution chain step-by-step and instantly terminate a run if it veers off course.

During the initial public preview phase, Google is heavily subsidizing the platform: developers will only be billed for standard Gemini 3.5 Flash token consumption, while the underlying virtual machine compute infrastructure is provided entirely free of charge.

Normally, when we use the Gemini API, the AI ​​only responds with text, and we have to write scripts on n8n to execute that text ourselves. However, Google's Managed Agents system provides the AI ​​with a dedicated backend computer (sandbox VM). When we send a command, the AI ​​writes and runs test code on that virtual machine, checks for bugs independently, and then sends the final, clean result back to us via a single API. It's like hiring a senior engineer to do the work for you at an affordable price.

The issue of defining Egress Network Policy is a key selling point for the enterprise side. In real-world operations, bank accounts and customer databases are sensitive. If the AI ​​experiences hallucinations or is subjected to cyberattacks through prompt injection, it could secretly send company data to external servers. Strict domain locking (e.g., allowing communication only to GitHub and internal cloud) alleviates concerns for security managers and encourages organizations to more readily adopt AI-powered code execution backend systems.

Google's decision to not charge for running virtual machines, but only for Gemini 3.5 Flash tokens, is a very clever move to attract developers away from OpenAI and Anthropic. Gemini 3.5 Flash is already known for its blazing-fast processing speeds (over 300 tokens/sec) and very low cost. Running this high-speed, low-cost duo on an agent simulation system allows developers to write iterative scripting bots thousands of times at a low cost. This effectively creates developer lock-in, familiarizing users with the Antigravity and Google Cloud ecosystem before fully charging for VMs in the future.

 

Samsung DRAM Monopoly Threatened by Massive 18-Day Strike Over Bonus Formulas. 

 

Source: Google Blog