How well does agent-sandbox work for coding agents on Kubernetes?

November 19, 2026 5 answers
Asked by 6x+gobletratchet_418

I'm evaluating agent-sandbox on EKS for short-lived environments where coding agents clone repositories, run builds and tests, and execute shell commands.

How do you choose an isolation runtime, and do warm pools make startup responsive enough for interactive use? I'm also interested in egress controls, dependency caching, and cleanup gotchas.

Answers

2w!ploverstencil_846

I'd separate the sandbox lifecycle from the security boundary. Agent-sandbox gives you a Kubernetes sandbox abstraction, but the isolation you get still depends on the pod configuration and runtime. A plain container runtime, gVisor and Kata are different choices, not interchangeable security settings.

Pick based on what the agent can execute and what you're trying to protect. Then test the actual build and test commands under each candidate. The discussion doesn't provide enough comparable measurements to say what the performance penalty will be for your workload.

Reply to this answer
_a63!torvexa.5

For latency, measure the whole path from a request arriving to the agent being able to run a command. Pod readiness alone may leave out repository checkout, dependency setup and anything else your harness needs.

Compare cold launches with requests served from a warm pool, including bursts that exhaust the pool. Record image size, node capacity and pool replenishment alongside the timings. A small image can help, but there's no useful startup number in the discussion to use as an EKS benchmark.

Reply to this answer
5q!pluvano_627

Network policy deserves its own test plan. Try the allowed paths, such as source control and package registries, and also verify that a sandbox cannot reach internal services or the cloud metadata endpoint. Don't assume a policy is doing what you intended without testing from inside the sandbox.

For dependency caching, a shared package proxy is worth evaluating before sharing writable workspace volumes between sessions. Keep the workspace itself scoped to a session, and consider whether cached artifacts could be poisoned by an untrusted run.

Reply to this answer
_r87+plendixa.2

Don't overlook the boring lifecycle stuff. Set CPU, memory and storage limits, put a timeout on sessions, and check what remains after normal completion, failure and forced deletion. Include sandbox resources, pods and any persistent volumes in that cleanup check.

The source discussion raises orphaned volumes and quota exhaustion as concerns, but doesn't establish how often they happen with agent-sandbox. They're worth testing rather than assuming either way.

Reply to this answer
_h74!zeldravo.6

There are other routes if the abstraction doesn't fit. A pod per session with a conventional runtime is simpler to evaluate, while Kata, Nitro-based sandboxes or a separate sandbox orchestration layer are alternatives to investigate for stronger isolation needs.

Those aren't equivalent setups, though. Compare the operational work and isolation boundary as well as startup time. The examples in the discussion don't establish a clear winner or provide enough detail for a production recommendation.

Reply to this answer

Add your answer

Answers are reviewed before publication.

Related Discussions