1. How Parmana works

Parmana sits between an AI agent and the systems it acts on. The agent never holds the credentials of those systems. It asks Parmana to perform an action, Parmana decides under a policy a person approved, and only an approved action is released to the system. Every decision, approved or refused, leaves a signed record anyone can check.

One request, in order

#	What happens	Where to read more
1	The agent asks which policy governs the action: `GET /policies/in-effect?capability=...`. The answer names the policy, its version in effect, and the signals a request must carry.	Chapter 3
2	The agent sends a Business Transaction to `POST /execute`: who is acting (authority), for what purpose (authorization), what to do (intent: action, target, parameters), the policy it declares, and its signals.	Chapter 3
3	Parmana checks the caller’s key may use the action, and that the declared policy is the one bound to the action at the version in effect.	Policies and the decision
4	Signals bound to the request are checked against it, and an approval signal counts as true only with a valid signed approval from a trusted person for that action and resource.	Human approval
5	The policy’s rules run in order; the first that matches decides. The same inputs always give the same decision.	Chapter 4
6	Refused: Parmana answers with the reason and stores a signed Refusal Record. Nothing is released.	Refusal Records
7	Approved: Parmana signs an execution authorization, single use, time bounded, and bound to the exact content approved.	Execution authorization
8	Before release it stores a signed Execution Intent: exactly what is about to be released.	Execution Intents
9	The gateway verifies the authorization again (signature, expiry, content, single use, the policy, the signals and the approval) and releases the action to the connector for that action.	The gateway
10	The connector acts with a credential only it can use, for this one execution.	Credential isolation
11	Parmana signs an Execution Trust Record: the transaction, the decision, the authorization and the connector’s evidence.	Execution Trust Records
12	Anyone with the public key can verify the record offline, without trusting the server.	Chapter 7

The ideas that matter

No action without a person. Every policy must require a signed human approval, reads included. A server refuses to load a policy that could approve without one. Plan for the first request on a new resource to be refused until a person signs.
The agent cannot authorize itself. Facts the agent declares can refuse an action, never authorize one on their own. Only a signed approval, or a fact the server checks for itself, can make an approve rule match.
Parmana releases the action; the agent does not. An agent that performs the action itself after asking Parmana defeats the control. The agent only reads the answer.
Changes need two people. A new policy version, a new approver key and a new external connector each need one person to propose and a different person to approve with a step up signature. No deploy is needed.
Everything is evidence. Approvals, refusals, intents and results are signed and kept. A refusal is as provable as an approval.

The words used in this book

Word	Meaning
Capability, action	What the agent wants done, `namespace:verb`, for example `paytm:refund`.
Business Transaction	One request to act, with its identifiers, intent, policy and signals.
Signal	A fact in the request that a policy reads, such as `refundAmount` or `managerApproved`.
Approval signal	A signal that counts as true only with a signed approval (`signals.approvalArtifact`).
Policy in effect	The version of the policy bound to an action that was most recently approved.
Maker, checker	The person who proposes a change, and the different person who approves it.
Connector	The part that performs an approved action on a system: built in (HubSpot, GitHub, Paytm, Slack) or external (your endpoint).

Build with Parmana

Reference

1. How Parmana works

One request, in order

The ideas that matter

The words used in this book

​One request, in order

​The ideas that matter

​The words used in this book

One request, in order

The ideas that matter

The words used in this book