Identity
Know which exact agent installation initiated the request.
Early · built in public
Give agents economic agency. Keep humans in control.
AG Pay is the payment control plane for AI agents—distinct identities, explicit permissions, and attributable purchases, without putting card details in model context.
Human-supervised by default Model-safe payment boundary Designed in public
A3
REQUESTED BYResearch agent 03
PAIREDMERCHANTData API · Pro plan
TOTAL$29.00/ month
AGENT RATIONALE
Required to validate supplier data for the current research task.
POLICY TRACE
IDENTITYVerified
POLICYReview required
DECISIONWaiting for human
PAYMENT BOUNDARYSecrets stay outside the model.
01 / 05The problem
Intelligence can choose. Trust must authorize.
Agents can research products, compare providers, and recommend subscriptions. But the moment they need to spend, capability runs into a trust problem.
“A card in a prompt is not a permission system.”
Economic action needs identity, limits, human judgment, and a record that survives the agent session.
Authority
Assign approved payment methods and review rules per agent.
Context
See the merchant, rationale, amount, and recurrence before deciding.
Accountability
Keep the agent and payment method attached to every outcome.
02 / 05Why now
Agents are moving from answers to actions.
Payments are the hard boundary between what an agent can reason about and what people can responsibly let it do. The standards for identity, oversight, and accountability are being shaped now.
01Research
Understands intent
02Decide
Compares options
03Request
Proposes a purchase
04Authorize
Applies human policy
05Act
Records the outcome
The next leap in agent capability is not unlimited autonomy. It is governable autonomy.
03 / 05The control plane
One control plane. Every decision attributable.
AG Pay links each proposal to an authenticated agent, an owner-approved payment method, and a human-defined policy.
01
Pair an agent
Give each runtime a distinct, revocable identity through a one-time pairing flow.
02
Set its boundaries
Assign only approved payment methods and choose when the agent must ask for review.
03
Review with context
Understand what the agent wants, where, why, and for how much. Approve or cancel.
04
Keep the history
Trace purchases and subscriptions back to the originating agent and payment method.
04 / 05The hard boundary
Useful to agents. Invisible to secrets.
The model gets the context it needs to make a request—not the credentials needed to move money. Raw card details stay out of model context and persistent AG Pay storage.
Separate human and agent identities Provider references instead of stored card numbers Human review is the fail-safe default
MODEL CONTEXT
Purchase intentitem · reason · amount
Sanitized outcomeapproved · failed · complete
AG PAY CONTROL PLANENO RAW CARD DATA
TRUSTED BOUNDARY
Payment provideropaque reference · safe metadata
Human policyidentity · rules · decision
IMPLEMENTED PROTOTYPESupervised autonomy, agent pairing, per-agent review rules, payment-method assignment, approvals, and attributable history. Broader live payment execution still requires provider integrations, compliance review, and production hardening.
05 / 05Why open source
Trust needs daylight.
Payments, identity, and agent autonomy are too consequential for opaque defaults. We believe this infrastructure should be built in the open—where the community can inspect the threat model, challenge approval boundaries, and help shape what comes next.
AG Pay is early. That is the invitation.
01
Challenge the threat model
Payments + security
02Shape policy primitives
Agents + protocols
03Improve the approval experience
Product + design
04Build the missing adapters
Infrastructure + tools
THE PAYMENT LAYER FOR AGENTS