Simulated. No real agent, store, bank, or standard takes part. The hashes are real SHA-256 in your browser. The signatures are fake.
1. The purchase
2. The dispute desk
What exists today, per public reports
| Format | What it covers | What it leaves out |
|---|
The bundle is a proposal. It does not replace these. It links records from them.
Try this
- Press Play. Watch the chat until the user opens a dispute.
- Read the desk with Without bundle. Count the formats and shared ids.
- Press With bundle. Tick the box that edits the merchant record. See the chain break.
Pain point
Five parties cannot agree on what an agent was allowed to buy when something goes wrong.
Problem statement
Each of five parties keeps its own record in its own format. A dispute team must collect and compare them all.
What sparked it
Per public reports on 2026-10-22, US banks detect and delete third-party AI agents in customer accounts. A PNC executive asked for authentication and liability law.
Why this approach
The mandate, cart, order, and payment link into one signed bundle. A dispute desk reads it in one view.
Hypothesis tested
We believed that one linked bundle lets a dispute desk read a case in one view. If it holds, card networks could ask for a bundle on the lighter dispute path.
How to try it
- Open the Prototype tab. Press Play, or Step.
- Read the desk under Without bundle.
- Press With bundle and read the verdict table.
- Edit the merchant record. Find the block that breaks.
What this does not do
- It uses no real agent, bank, or card network. All data is fake.
- It signs nothing. Only the hashes are real.
- The split of loss is an illustration.
Sources
- Banks and AI agents (Reuters, citing Axios):
https://www.pymnts.com/news/banking/2026/banks-crack-down-on-unauthorized-ai-agents-operating-inside-customer-accounts/ - American Express framework, 2026-10-13:
https://techcrunch.com/2026/10/13/amex-introduces-a-framework-to-prevent-ai-agents-from-making-unauthorized-purchases/ - PACT, 2026-10-06:
https://decagon.ai/blog/introducing-the-personal-agent-consent-trust-protocol-pact
Notes
- Who would pay: dispute teams at card issuers, acquirers, and processors.
- First customer: a payment processor with a dispute product.
- Closest alternatives: AP2, the card networks, and PACT, each in part.
- Weak spot: no confirmed agent-purchase loss exists yet.
- Feedback: would your team accept one new bundle, or only a translator?
Traces
| Step | Prompt | Model | Tool | Date |
|---|
Cost
| Step | Model | Input tokens | Output tokens | Cost (USD) |
|---|---|---|---|---|
| Total | ||||
Measured: the author runs used 120,733 tokens in total. Estimated: the split into input and output tokens (90 percent input, 10 percent output) and the dollar cost, at $2 input and $10 output per million tokens for Sonnet 5.5 (about $0.34). The run reported one total only. The lead session's planning, research, and publishing are not included. The prompts in the Traces section are listed in the order they were given.