Why Parmana exists
- Your agent’s claim is not a business fact
- Prompt guardrails are advice. Authorization is a gate.
- Agents will rewrite themselves. Their authority should not move when they do.
The decision
- Policies as data: same input, same decision
- Every action needs a signed human approval, reads included
- Giving an agent authority: grants with limits and an end date
- Per-request limits are not enough
The release
- The gateway: why your agent should never hold the keys
- Signed, single-use, short-lived: the execution authorization
- Approved what, exactly? Binding approval to content
- Write it down before you do it
- Proving the gateway is the gateway
The evidence
- An approval you can verify without trusting me
- Proving what didn’t happen
- Who asked? The caller audit trail
- Replay: what it can and can’t re-check
- Verify it yourself: offline and in the browser
- Choosing signatures: Ed25519, post-quantum, and keys in KMS
Governance and operations
- Maker checker for one person
- Governing the rules
- Rotating every key in production in one day
- Running Parmana yourself, even with no internet
Connecting agents and systems
- Connecting an agent
- Connecting any system
- What a refund agent needs before it touches a payment gateway
Proof and honesty
- Testing agent controls in production without moving real money
- What “done” means: deployed, but not yet deciding anything
- What Parmana can’t stop
- Mapping to regulation without overclaiming
- Where Parmana fits among agent controls