Projects / Agent infrastructure
A safe place for agents to run code
Agent infrastructure AI product client
✦ Result
10k
sandboxed agent runs a day
Isolated sandboxes where AI agents can run code, browse and use tools, started in milliseconds and thrown away after.
Before: 0 agents allowed near real systems
The problem
The client’s agents needed to run code and use tools, but nobody would let them near real systems. Every experiment meant a hand-built VM.
What we built
A sandbox service: each agent run gets its own isolated machine with its own files and network rules, starts in under 300 ms, and is destroyed when the run ends. A gateway decides which tools and secrets each run can use.
Results
Agents now run 10,000 jobs a day in sandboxes, and security signed off on letting them touch staging.
How it was built
- Firecracker microVMs
- Network policies
- Secrets broker
- Per-run file systems
- Tool gateway
- Tracing
- Kubernetes
- Terraform
- Weeks 1–2: threat model and design
- Weeks 3–6: sandbox runtime and gateway
- Weeks 7–8: red-team and load tests
- Week 9: live
- 1 lead engineer
- 1 DevOps engineer
- 1 security reviewer
FAQs
Why not run agents in a container?
Containers share a kernel. For untrusted code we wanted a harder wall, and still a fast start.
Can agents reach the internet?
Only what the policy allows, per run. Everything is logged.
Got a process like this? 20 minutes, and you'll know if it's worth building.
Book a call