@RoundtableSpace Even if the security protocols within the agent are bolstered, it is wise to have some external guardrails and an auditing mechanism. We built AgentPays (https://t.co/90RMLJbt8C) to do this for payment agents. Control layers and human oversight will still prove invaluable
Agents will still use standard payment methods and bank accounts. You talk about friction in the economy. Yet, the friction of getting every company to accept crypto/blockchain payments/systems of record is far more friction than giving an AI traditional payment methods.
AI agents are an extension of the user, so the user can create an account and give the agent access to do certain things. Similar to how companies give employees company cards.
Yes, guardrails will be necessary and that’s why I created https://t.co/iDi8ZHItrp. A solution that allows agents to buy anything now (not just accepted at <1% of places like crypto) and stops the agent from spending on things it shouldn’t.
I think right now you may need a mix of both. Better to be safe than sorry. I would recommend setting a low auto-approve threshold for only certain merchants at first. See what your agent requests and if it does well. Then, you can raise auto-approve threshold and expand approved merchants list.
Also, if you have more than one agent purchasing, you may want to have cross-agent limits and rules. Or even a list of what the other agents bought so you don't accidentally buy the same thing twice.
Handing your AI agent your credit card is like handing your credit card to an elementary school student.
You’d worry the agent would buy things you don’t want, accidentally buy the same thing 10 times because it wasn’t sure it went through, or get defrauded by a malicious website.
In many ways, this is what is holding back the agent economy.
An extra benefit of forcing the agent to make the purchase request to AgentPays is that it leaves a clear audit trail of what it bought from where and why which can then be used to improve agent purchasing in the future.
If this sounds like something that would be valuable to you or your business please check it out: https://t.co/iDi8ZHItrp and would love any feedback.
Right now, we are still new but we have an enterprise grade system coming soon so if you have any interest in the enterprise version, please let me know because that will help us prioritize the right features.
This problem is why I created AgentPays.
AgentPays gives agents a way to pay with far less risk. Users can set per agent rules including spending limits, merchant allowed/disallowed list, and auto-approve thresholds.
Then, when an agent wants to make a purchase, it makes a call to the AgentPays MCP server with what it wants to buy, how much it costs, from what website url, and why.
If it meets the rules and is under the auto-approve threshold, it’s automatically approved and the agent gets sent a one time link with a single use virtual card payment details to complete the check out.
If the rules don't allow the transaction, there is no virtual card created and the agent gets told it was not approved.
If it doesn’t break a rule but is over the auto-approve threshold, the user gets an approval email.
This way, if something were to go wrong, the blast radius would be limited to the funds on the one card (which are only available for 20 mins which decreases risk even further).