Governance & security

Agents read freely. Changes wait for a person.

Entropix is built so an agent can be useful without being trusted with everything: it reads, proposes and explains — and a member of your team approves anything that changes the world.

STEP 1
The agent proposes

An asks-first tool is offered in a goal or in chat. In a goal, calling it pauses the goal — one pending action at a time. In chat, the agent asks in the conversation.

STEP 2
A person approves

In a goal, a member who may decide sees the exact arguments. In chat, a typed yes or no answers. No answer in 24 hours, and a goal's action expires.

STEP 3
It runs once, on record

Under the goal's lease, exactly as approved. An unknown outcome is never re-run. Everything lands in the Actions log.

Actions log · platform-agent
ToolArgumentsDecided byOutcome
patch_deployment checkout-api memory=768Mi Priya ran
create_issue OPS “Raise OOM alert” Sam ran
scale_deployment checkout-api 4 → 6 — waiting
delete_pod payments-7f9c Alex declined
rollout_restart search-api — expired
Built in, not bolted on

The rules the portal holds itself to.

Code stays on your machines

Only local sessions change code. The portal never writes to a repository.

Git, read-only

SSH with your organisation's keys, fetch only, strict host keys, no hooks. The key never leaves the reader.

Organisations kept apart

Every record carries its organisation. Another organisation's record simply does not exist for you.

Roles and a permission table

Every change names the permission it needs, checked before anything else happens.

People and agents never share keys

People sign in with a session; coding sessions use a personal access token. Neither works for the other.

A person approves what a model writes

Nothing a model drafts about a repository, rule or tool reaches an agent until someone approves it.

An audit trail of every step

Every action and agent step is recorded, so you can see what an agent did and who approved it.

Coach scores

Each agent is scored on repository knowledge, system understanding, memory, tools and honesty, and improved through changes a maintainer applies.

Questions security teams ask

Does Entropix see our source code?

It reads the repositories you add — to scan them for an agent and to review merge requests. It never writes to them.

Which model does it use?

The one you point it at. With no model configured, the portal calls none.

Can an agent change production on its own?

Not through a tool marked asks first. Those run only after a member approves the exact call, once.

Where is our data hosted?

Hosted in Sydney, Australia (AWS ap-southeast-2).

What is in place, and what is not yet: the Security page. Security questions: Contact us and choose Security.