Ever tried adding a second item to your cart and wondered why the checkout never moves? This video reveals why AI agents struggle with those tiny shopping details.
The Simple Version
Think of a shopping agent like a cat trying to fetch a toy: it looks cute but gets stuck on tiny rules.
How It Actually Works
Agents need to talk
Agents use A2A to exchange messages with merchant agents.
Tool access
MCP gives agents permission to click, read, or buy.
Commerce primitives
ACP from OpenAI and UCP from Google are the two recipes.
Payment mandates
AP2 carries authorizing party, spend ceiling, and revocation URL.
Technical Deep Dive
Agnaf Prio demonstrates the full loop on Cerebras at 3,000 tokens per second.
An inspector shows every call and checkout state transition in real time.
Product feeds replace slow catalog searches because m merchants times n products does not scale.
He also runs a live demo where his orange tabby Ginny haggles a discount code.
Why This Matters
When agents can shop reliably, retailers gain 45% of sessions that actually convert.
Merchants avoid fraud alarms by using structured payment mandates instead of raw browser automation.
Fast token rates mean checkout flows can keep up with real user expectations.
Key Takeaways
- Agentic shopping fails on tiny SKU details.
- MCP grants agents safe tool access.
- A2A enables agent‑to‑agent communication.
- ACP and UCP are competing commerce primitives.
- AP2 adds secure payment controls.
- Product feeds solve catalog scaling.
- Evals focus on behavior, latency, and protocol compliance.