In his conversation with Zach Bruggeman, Gabriel Benmergui asked how much of Ramp's infrastructure runs inside an Inspect sandbox. Zach's answer gets at a constraint founders face when adopting coding agents: an agent needs a working environment before it can check what it changed.
Inspect is Ramp's background coding agent. In the clip, Zach describes running the part of the application that the task needs and preparing virtual-machine images ahead of time. The goal is to let the agent get to work without waiting for dependencies to install.
If you're hiring engineers in Latin America, ask candidates how they'd make your application easy to start and give an agent useful tests.
Run the surface the task needs
Zach says Inspect's sandbox runs the specific surface the agent is working on. A frontend task gets the frontend running. A backend task gets the backend. He doesn't describe starting the entire Ramp infrastructure for each change.
Ask an engineer to start the frontend from a fresh environment and run its checks. Watch which steps depend on knowledge that lives in someone's head. An undocumented command or missing dependency is something the team needs to resolve before it can expect an agent to work independently.
Scoping also makes failures easier to investigate. If the task concerns a frontend change, you want the checks to tell you whether that change works. You shouldn't have to diagnose an unrelated service before you can evaluate it. Keeping those boundaries usable takes engineering attention as the application grows.
Full-stack verification needs both sides
Zach describes a planned improvement. They want more full-stack flows: for a task that changes the backend and frontend, the agent should be able to run both and verify the result.
"we also want to do more of these kind of like fullstack flows"
That is a direction he describes, not a claim that Inspect already verifies every end-to-end change. Gabriel follows up with the prerequisite: each surface has to run on its own and give you a useful signal.
For your team, a frontend check and a backend check answer different questions from a check of the complete flow. A passing test on either side doesn't establish that they work together. Engineers need to decide which checks fit the change and make the required services available. They also need to distinguish a broken environment from a failed assertion about the product.
Our agent-ready discussion considers how agents use a product's interfaces. The sandbox question concerns the engineers building those interfaces: can they run enough of the application to test their work?
Prepare the image before the task starts
Zach points out that the technology to spin up virtual machines quickly already exists. He describes Ramp putting work into pre-building images with the software the agent needs installed. The intended result is that the machines can:
"start up very quickly and get to work"
The distinction is between starting a machine and having an application ready to use. If the agent spends its session installing dependencies, a fast VM launch has only solved part of the problem.
An image also needs maintenance as your project changes. Ask who updates it when the team changes dependencies and how they check that a fresh sandbox still starts. A reproducible environment lets another engineer investigate a failure without reconstructing the original setup by hand.
Hire for the work around the agent
You can use a small task to evaluate this in an interview. Ask a candidate to make a change with an agent, explain the environment it needs, and show how they checked the result. Follow the work through a failure. Pay attention to whether they investigate the cause or accept the agent's explanation.
Silver.dev's AI adoption kickstart and AI adoption interview guide, both in Spanish, explain how we prepare candidates and evaluate their work with agents. Our agents versus product discussion covers the product judgment that still belongs with your team.
If you're hiring engineers in Latin America, ask candidates to leave behind commands a teammate can run and evidence that the change works.
Watch the clip from Gabriel's conversation with Zach Bruggeman on this page and open the transcript to follow Zach's explanation.