hermes atlas
217·repos hermes·v0.19.0 ★ star this repo

Guardrails before it touches anything real

“I want approval gates and sandboxing so my agent can't do something destructive on its own”

Almost every guardrail story starts the same way — the agent did something confidently wrong, fast, and the person went looking for a brake. Layer these from outside in: a sandbox that limits the blast radius regardless of what the agent believes, a runtime harness for the credentials it can see, then policy that decides which actions require a human. The trust-boundary fork is the alternative to bolting the middle layers on, and it is a real fork, so treat it as an either/or rather than an addition. Adversarial review is narrower — it protects the code you ship, not the system the agent runs on.

Real posts from the Hermes community, mirrored from Nous Research's user stories corpus.