How Antithesis Built a Workflow for AI Agent Testing Using exe
Key Takeaways
- ▸Antithesis solved the sandbox paradox for AI agents—providing sufficient autonomy and root access without over-constraining capability or introducing reckless risk
- ▸exe's per-user resource pooling and shared VM quota eliminate cost friction, allowing developers to treat sandboxes as ephemeral and disposable rather than precious resources
- ▸Root-access VMs are essential for agentic testing; restrictive sandboxes block AI agents from performing real diagnostic and testing work
Summary
Antithesis, an autonomous software testing platform, has developed an innovative approach to deploying AI agents for testing by solving a fundamental challenge: balancing agent autonomy with safety. Too much constraint and agents cannot accomplish real work; too much freedom and they risk destabilizing systems. Carl Sverre, a software engineer at Antithesis, discovered that exe—a VM-based sandbox platform—provides the ideal middle ground for agentic testing workflows.
The key advantage of exe is its architectural and pricing approach. Unlike most VM platforms that charge per machine regardless of utilization, exe uses a per-user pricing model ($25/month per team member) with access to up to 50 VMs drawing from a shared resource pool. This eliminates the financial friction that plagues sandbox solutions and allows developers to freely spin up ephemeral testing environments. Additionally, exe provides full root access to VMs—critical for allowing AI agents to execute complex testing tasks that would be blocked by more restrictive sandboxes—while avoiding hibernation delays that plague other platforms.
Antithesis has integrated exe into two workflows. For rapid bug-catching on open-source projects, developers spin up temporary VMs, invoke Claude via the platform, and then discard the machine. For persistent development work like building Snouty (Antithesis's CLI tool), teams maintain long-running VMs with Claude remote-control access from web and mobile. When Sverre identified a security concern with exe's VM sharing model, his direct collaboration with exe engineers led to a quick product fix, demonstrating the close partnership between the platforms.
- Antithesis influenced exe's product roadmap, demonstrating the value of close collaboration between testing platforms and infrastructure providers
