Capability Manifest
The up-front declaration of what an agent is permitted to do, stored per organization rather than implied by a prompt.
A capability manifest is the declaration of what an agent is permitted to do, written before the agent acts and stored at the organization level rather than implied by whatever its prompt happens to say. It is the difference between an agent whose limits are discoverable and one whose limits are discovered.
The distinction is practical. A prompt is instruction: it describes intent, it can be overridden by context, and it is not legible to anyone who is not reading the prompt. A manifest is declaration: it states the permitted action set independently of any single run, it survives changes to the model or the instructions, and a compliance reviewer can read it without reading a prompt at all.
Manifests are the basis for governing an agent before it acts rather than reviewing it afterward, which matters because agents act faster than humans notice. In FlashyOS, per-agent capabilities are declared and stored per organization, and are paired with auto-accept policies that state by category what agents may do without asking. Actions outside the declared set do not silently proceed; they route to a decision with an impact level and a named human resolver.
Common questions
Why is a capability manifest better than describing limits in a prompt?
A prompt is instruction that context can override and that only prompt readers can inspect. A manifest is a stored declaration tied to the organization, so it survives model and prompt changes and can be reviewed by people who never see the prompt.
What happens when an agent tries something outside its manifest?
It should not proceed silently. In FlashyOS the action becomes a decision carrying an impact level and a status, routed to a named human resolver, with the whole sequence written to the append-only session log.