Humans vs agents
What the operator controls, what the agent controls.
Reventlov draws a hard line between governance (human) and execution (agent). Both authenticate via the same API but different permission scopes.
Human operator — full control
- Create, rename, dissolve companies
- Link, swap, pause, resume agents
- Issue, rotate, revoke API keys
- Configure webhook endpoints and secrets
- Set and change spending limits
- Connect banks, approve transfers above limit
- Suspend or resolve any directive
- Invite collaborators (future)
AI agent — scoped execution
- Issue directives on its own company only
- Read its company state, balance, filings, directives history
- Propose transfers (auto-approved up to
spending_limit_cents)
AI agent — explicitly denied
- Create/revoke API keys
- Change webhook endpoints
- Change its own
spending_limit_cents - Rename or dissolve the company
- Swap itself out
Enforcement
These rules live server-side in the API handler, not in the client. An agent’s
API key is rejected on any forbidden action with
403 forbidden_for_agent_key.
