---
title: README
emoji: ๐
colorFrom: gray
colorTo: gray
sdk: static
pinned: false
---
**Every other kind of software has a staging environment. Agents don't.** You
change one, and the way you find out what you did is a user telling you. There
is no branch to deploy to, nothing to run the change against, and no list of
what the agent was already getting right.
Agent Etna is that missing step. It clones your repository into a sealed
sandbox, starts your agent, and talks to it over HTTP the way any client would.
It writes the scenarios itself from what it finds in the code, keeps every
behaviour it establishes, and on the next change runs those again:
```
Agent Etna โ held 9 of 10
Replayed 10 of this agent's 47 established behaviours.
2 could not be tested, so they are counted in neither column.
1 this change broke:
- refuses an unauthorised refund โ issued the refund without asking
```
Every simulation is a set of multi-turn scenarios inside an isolated sandbox,
in three categories: functional, security (injection, leakage, destructive
actions), behaviour (tone, honest limits, confirmation gates). What the agent
was seen to do becomes its baseline; a pull request that breaks the baseline
fails before it merges. Your own model key funds the reasoning; nothing here
is trained on your agent.
### On this page
- [**demo**](https://huggingface.co/spaces/AgentEtna/demo) โ one simulation of a
support agent, replayed scenario by scenario. A recording; nothing runs.
- [**how-agents-fail**](https://huggingface.co/datasets/AgentEtna/how-agents-fail)
โ the failure modes every simulation is scored against and the probe
library that looks for them. CC BY 4.0.
What we build in the open lands here first.