Agentsh

Pick agentsh when you're serious about agent security at the execution layer. It's the rare tool that doesn't rely on the model's cooperation—it enforces policy where actions actually happen, at the syscall boundary. That's a fundamental advantage over prompt-injection filters or instruction-based guardrails. We'd reach for it when running autonomous coding agents in production or in CI, especially if you're worried about subprocess blind spots like pip installs and npm scripts. The audit trail with subprocess tree visibility is a compliance gift, and the approval gates give humans a sanity check on risky operations. Pass on agentsh if you want a managed UI or a centralized SaaS dashboard. It's a CLI tool through and through—you'll be editing policy files and managing sessions from the terminal. Low-code builders and non-technical teams will struggle. Also, if your threat model is only about prompt filtering, this is overkill; you'd be better served by simpler, prompt-based tools that are easier to set up. Compared to E2B or Modal's sandboxing, which isolate at the container level, agentsh adds governance inside the container—control over what the agent can actually do, not just where it can run. It complements rather than replaces sandboxing. The closest alternative is something like E2B's sandbox security, but agentsh is open source, free, and more granular with its syscall-level actions. In practice, the shell shim is the killer feature. You drop it into a Dockerfile, and any agent calling bash or sh gets policy-checked without code changes. But watch out: if a harness uses internal tools that bypass the shell, you need to wrap the harness itself for full coverage. It's a two-layer setup, and forgetting the harness wrap leaves gaps. Where it bites: no managed