How to Scale AI Agents with Secure Sandboxes

TL;DR

Give each agent its own sandboxed computer so it can work independently without colliding with your local environment or other agents. Placing the application, agent configurations, developer workflow, and an internal orchestrator inside durable sandboxes enables isolation, parallel scale, and greater autonomy, including Best of N workflows that produce multiple solutions for comparison.

Transcript

What's up engineers? Indy Devdan here. It's time to focus. This is going to be a big video and a lot of engineers are going to miss the leverage I'm going to share in this one. I don't want you to be one of them. Your software factory can unlock an unprecedented level of engineering results. But there's a massive roadblock every agentic engineer ru... Read More

Key Insights

• Agent sandboxes are dedicated developer computers for agents, providing an isolated place where code, tools, applications, and workflows can run. Giving each agent an entire environment prevents the agent from occupying a small corner of the engineer's own machine and reduces interference between concurrent jobs.
• The three primary benefits of agent sandboxes are isolation, scale, and autonomy. Isolation separates workloads, scale permits multiple environments to run concurrently, and autonomy gives each agent control of the resources it needs to pursue an assigned outcome without continuous human supervision.
• Best of N is a workflow that assigns the same problem to several independent agent configurations. The demonstration launches five sandboxes using Default, Frontier, Deepest Seek, Open Weights, and Top Speed configurations, then exposes their separate applications and software factories for comparison.
• A software factory is more than a model responding to a prompt. It combines a selected model with surrounding code, an agent harness, available tools, orchestration, and a developer workflow that moves work through the software development life cycle.
• The top-level orchestrator runs outside the sandboxes and creates the parallel workflow. It provisions separate environments and places a complete software factory inside each one, allowing a single prompt to initiate multiple independent attempts at the same application task.
• Each sandbox contains both an application layer and an agentic layer. The application shows the resulting product, while the software-factory interface provides observability into the models, agent harnesses, tools, and workflow operating inside that environment.
• Human involvement can become the bottleneck when every agent action requires attention or approval. The proposed approach moves engineers outside routine execution loops, while recognizing that hands-on involvement remains useful when engineers are building or refining the system that creates other systems.
• Model selection is presented as only one part of agentic engineering. The more important design question is which model should run in which environment with which supporting code, tools, and harness, because the complete configuration determines how effectively a software factory performs its job.

Questions & Answers

Q: What is an agent sandbox for software development?

An agent sandbox is a dedicated developer computer where an AI agent can run code, tools, applications, and development workflows. Instead of receiving a small portion of an engineer's local machine, the agent owns an isolated environment. The source describes these sandboxes as durable, fast, secure, and shareable environments that support isolation, scale, and autonomous execution.

Q: Why do software factories need agent sandboxes?

Software factories need agent sandboxes because local machines, shared containers, and excessive reliance on CI/CD can restrict independent execution. A sandbox gives each factory its own environment, reducing collisions with the engineer and other agents. It also lets the complete factory, including the application, agent configurations, tools, workflow, and orchestrator, operate together without constant human involvement.

Q: How do agent sandboxes improve isolation and security?

Agent sandboxes improve isolation by placing each agent and its code in a separate computer environment. Parallel workers do not need to share a small corner of the engineer's device, so their commands and resources are less likely to interfere with one another or with the engineer's work. The transcript connects this stronger isolation with safer, more autonomous agent operation.

Q: How does a Best of N agent workflow work?

A Best of N workflow sends the same problem to multiple independent agent configurations and compares the resulting implementations. In the demonstration, five configurations named Default, Frontier, Deepest Seek, Open Weights, and Top Speed each receive their own sandbox. Every sandbox runs a full developer workflow against the same Inkwell redesign prompt, creating five distinct solutions from which one can be chosen.

Q: What runs inside each software-factory sandbox?

Each sandbox contains more than a single coding agent. It includes the application, agent configurations, supporting code, available tools, an AI developer workflow, and an orchestrator agent. The workflow follows the software development life cycle and ships work on the engineer's behalf. Separate application and factory interfaces also make the resulting product and internal agent configuration observable.

Q: Why is there an orchestrator outside and inside the sandbox?

The outside orchestrator coordinates the overall Best of N process by creating separate sandboxes and starting a software factory in each one. An internal orchestrator then manages the agents and development workflow within its assigned environment. This layered structure separates global coordination from local execution, allowing every sandbox to operate as an independent factory while remaining part of one larger workflow.

Q: Why can keeping a human in the agent loop limit scale?

Keeping a human inside every execution loop limits scale because the agents must repeatedly wait for attention, input, or supervision. Moving routine development work into self-contained sandboxes lets agents execute workflows without the engineer. The transcript still identifies an exception: hands-on interaction is appropriate when the engineer is building, testing, or refining the system that builds other systems.

Q: How should engineers choose models for a software factory?

Engineers should choose models as parts of complete configurations rather than treating model selection as the entire problem. The relevant question is which model should run where and with what surrounding code, harness, and tools. The demonstrated factories intentionally use different configurations, including frontier, workhorse, lightweight, open-weights, and speed-focused options, so their results can be compared on the same task.

Summary & Key Takeaways

• Agent sandboxes give software agents dedicated developer environments instead of forcing them to share a small portion of a human engineer's computer. This arrangement provides isolation, scale, and autonomy, allowing agents to own their environments, use the tools they need, and complete work without interfering with the engineer or one another.

• The demonstrated Best of N workflow launches five separate sandboxes, each containing a software factory that executes the same application redesign task. Every factory uses a different combination of models, agent harnesses, and tools, producing five implementations that can be observed through separate factory and application URLs before one result is selected.

• The broader software-factory approach treats autonomous agents as software components within established developer workflows. A top-level orchestrator creates and coordinates sandboxes, while each sandbox contains the application, agent configurations, development workflow, and another orchestrator. This structure shifts engineers away from supervising every action and toward designing systems that can deliver outcomes independently.