Notes from the factory Preview
Keeping secrets out of the context window
Anything an agent reads can end up in a log, a commit or someone else's prompt. Agents should get capabilities, never credentials.
Open a recent coding agent transcript and search it for sk-, ghp_, AKIA or password=. If you find anything, a credential has already left your control. It is in a transcript on someone’s disk, maybe in a provider’s logs, possibly in a commit.
It gets there the ordinary way. An agent needs to call an API, so someone pastes a key into the chat. Or the agent reads .env to see how the app is configured, or a failing command prints its own connection string. Nothing malicious happened. The context window is the wrong place for a secret, and agents read everything.
Our rule is short: agents get capabilities, never credentials. The agent calls a tool, the tool holds the key, and the agent gets the result. It never holds the key itself.
We follow that rule with Heka Vault, a secrets vault we host ourselves, so our own agents work without ever seeing a credential. The full post will show how we set it up, and what to do first on your own team: search your transcripts, stop agents from reading .env, and move every pasted key behind a tool.
This is a preview: a short note ahead of a full post on the same subject. The full post is coming; RSS carries it when it is out.