Getting AI agents to production faster – protocol, technology and ROI

Today, I explore the progress made in getting AI agents closer to production.

Google Agent2Payment (AP2) protocol

The A2P protocol announced by Google recently, solves the problem of letting an AI agent perform a payment. The payment systems have been designed to allow a human to pay, but how can we extend the systems to trust AI agents as well? A2P solves it by providing answers to three questions

Prove that a user gave the agent permission to buy something and pay for it.
Allow the merchant to confirm that the agent’s actions reflect the human’s intent.
If fraud happens, provide accountability

The way it works is that humans provide an ‘intent mandate‘ to the agent describing what it wants to buy and what are the conditions around it. The mandates are tamper-proof. When the agent shows your what you want, you sign a ‘cart intent‘ which allows the agent to buy it on your behalf.

AP2 is supported by these payment systems: Adyen, American Express, Ant International, Coinbase, Etsy, Forter, Intuit, JCB, Mastercard, Mysten Labs, Paypal, Revolut, Salesforce, ServiceNow, UnionPay International, Worldpay

Progress towards making LLM deterministic

Thinking Machines lab published a paper that shows a technique to remove non deterministic behaviour of an LLM. So if you set the temperature to 0, and ask a question 100 times, you get the same response each time.

While floating-point arithmetic and concurrency are often blamed, the real reason is that batch size and the way GPU kernels process batches can cause outputs to change unpredictably. Variations in server load or parallel user requests result in different batch sizes, in turn causing slight numerical differences in computation, which propagate to different generated outputs. The article proposes making kernel operations “batch-invariant” as a solution for achieving deterministic LLM inference, even under changing server loads or concurrent requests.