A code of conduct for AI agents.
Agents that ship to real users carry real consequences. Agent Etna is built around a small, opinionated set of principles — and the product makes them defaults, not aspirations.
1. Reversibility before autonomy.
Our position: automation is only acceptable when humans can undo it instantly. Speed is not an excuse for trapping users in changes they didn't choose.
Where Agent Etna helps: every fix is one click away from being rolled back. Every parameter change, every prompt change, every model swap is versioned, signed, and reversible. The rollback itself is logged with a cryptographic signature so an audit trail survives the revert.
2. Provenance over speed.
Our position: every change to a production agent must be cryptographically traceable to its author. Anyone — auditor, teammate, future-you — should be able to prove what shipped and where it came from.
Where Agent Etna helps: every commit Agent Etna writes is cryptographically signed, tied to your instance. A public verification page renders the proof. Tampering is detectable. Speed is not bought at the cost of opacity.
3. Adversarial testing as duty.
Our position: agents should be stress-tested for harm before they meet users. "Happy path passes" is not enough; an agent that hasn't been probed for prompt injection, jailbreaks, or data exfiltration is not ready for production.
Where Agent Etna helps: every test run includes adversarial probes by default — prompt injection, jailbreaks, data exfiltration, and more. A green simulator means the agent held up against the kinds of inputs production sees — not just the ones we hoped for.
4. Transparency over confidence.
Our position: evaluations must surface weakness as prominently as strength. Headline averages that hide a 62/100 on safety behind an 89/100 mean are dishonest by design.
Where Agent Etna helps: every reply is scored on five dimensions — helpfulness, accuracy, safety, brand voice, conciseness — and the weakest dimension is surfaced explicitly. You see what's working and what isn't, before you ship.
5. Humans approve, Agent Etna deploys.
Our position: there is no autonomous-deployment toggle in a serious AI tool. The agent proposes; the human disposes. Removing the human from the loop removes accountability with it.
Where Agent Etna helps: every change is a normal GitHub PR labelled agent-etna, opened against your repo and ready for code review. No silent merges, no shadow commits, no automation that bypasses the people responsible for the system.
6. Your data, your keys, your control.
Our position: the simplest privacy guarantee is the one that doesn't require trust. Customer prompts, code, and conversations should never become training material without explicit consent — and where consent is given, the corpus must be inspectable and revocable.
Where Agent Etna helps: BYOK by default. Your prompts and your conversations stay attached to your own model account, not ours. The shadow-pair training pipeline is opt-in; the corpus is yours to inspect and revoke at any time.
7. Open about limitations.
Our position: we say what Agent Etna cannot do as plainly as what it can. There is no AGI claim, no autonomous-agent fantasy. Agent Etna is a control plane for the messy, human-supervised reality of shipping AI systems today — not a marketing brochure for one we hope arrives later.
Where Agent Etna helps: the documentation describes failure modes (hallucinations not caught by the rubric, model regressions across providers, edge cases the suite missed) explicitly. The product surfaces uncertainty in scoring rather than hiding it. Every claim on this site is something the system actually does.
See the code of conduct in action.
Connect a GitHub repo and run your first sandboxed, signed, transparently-scored fix. The principles aren't ornamental — they're how the product behaves by default.