Co-founder & CEO @BanzaOfficial
Dubai, United Arab Emirates
Joined January 2013
- Tweets2K
- Following3.9K
- Followers1.7K
- Likes13K
Pocket money.
That's the model I keep coming back to for agent wallets. A separate stablecoin wallet, topped up with a fixed amount every Monday, with no way to pull from your main account.
A bug, a bad prompt or a fake invoice can only spend what's left in that week's tin.
I'd have bet more info helps an agent bargain. Exact values helped some models and made others greedier. So the model and what you tell it need testing together. xitter.gq/rohanpaul_ai/status/21…
New Stanford paper shows that more information helps some LLM agents trade better and hurts others, depending on the model.
AI agents buying and selling in an auction were told exactly how much each side valued the goods, Claude Opus 5.5’s share of possible gains fell from 94% to 64%.
Multi-agent systems need rules on what each agent sees about others, such as exact prices or only rankings. The usual assumption is that more information helps.
They ran 10 frontier LLMs in trading, bargaining, and double auctions, changing only how precisely private values were shared, and scored gains against known utilities.
In trade, exact values cut GPT-6 Sol's gains from 98% to 91% of the maximum but raised Claude Sonnet 5's from 72% to 86%. In auctions, full disclosure inflated price demands and dropped Grok 4.7 from 97% to 62%.
When agents negotiate for you, test the model and what it gets told together, because more detail makes some models greedier.
500+ applicants and a bunch of bots. A screenshot plus a short critique is a smart filter.
The screenshot shows you actually used the product, and the critique shows whether you noticed anything real while you did. xitter.gq/jeff_weinstein/status/…
🤯 500+ people (and a bunch of bots...) emailed me applying to @stripe's agentic commerce team—I'm through 1/2 of them.
An ask: If you'd like to *dramatically* increase your chances, please send me any screenshot of using a Stripe product and write 3-5 sentences critiquing it.
Pieter Levels' 30,000 dollar sofa is about 4,500 if Claude Code designs it and you buy parts direct.
The daily version is dinner.
Same dish, often a different price on each app, and most of us open the usual one. The AI Twin at @BanzaOfficial checks the others before I order.
"Only 2 left."
Countdown timers.
A crossed out price.
A good shopping agent reads all three as plain data and moves on. Some retailers are now making holiday offers readable for bots as well as people. Curious how much of the old playbook survives a reader who can't be rushed.
Real users judging personal assistants. That's a score I trust more than a lab number.
Next time, add one task where the right move is doing nothing. xitter.gq/DesignArena/status/211…
Introducing LifeArena: the world’s first arena for personal assistants, as judged by real users.
As AI assistants become more capable, existing static benchmarks struggle to capture the creative, unthinkable requests of people in their everyday lives.
So, we offered select users the most popular personal assistants: @instinct, @Muse by @AIatMeta, Dots by @OpenAI, Grok Bot by @xai, and Gemini by @GoogleDeepMind.
The results are in: @Muse is #1... by a landslide.
Here's what users said:
Who gets to say no ? That's my test for every way an agent can pay.
Bot typing your card: the shop's fraud check
Virtual card: the platform
Network token: your bank
Stablecoin balance: only your wallet rules
On that last rail, your code is the bank and the fraud team.
A per purchase cap onchain stops the one big mistake. Ten small wrong buys that each fit under it go straight through. Per merchant and per day limits matter just as much. xitter.gq/hedera/status/21187283…
“AI agents can hallucinate and cannot always be trusted.”
@Lindsayjowalk @hashgraph explains why agentic commerce needs deterministic, on-chain spending limits and tamper-proof records, in an interview with @futuristdotai of @securitiesio 🔐
securities.io/lindsay-walker…
Visa gave 22 AI models 15 shopping tasks and three tries at each. All three tries came out right about a third of the time.
That matters more to me than the top score. An agent that picks the right laptop on Monday and the wrong one on Thursday still has me checking every order.
48 hours of agent launches. Agents that call, shop, check out and pay. The slow part for me was always the decision before all that. The AI Twin we build at @BanzaOfficial is for that moment. A short list, one clear pick, and it tells me why. I still make the final call.
Clean product data used to be an SEO chore. Now it's the sales pitch.
Same on the shopper side : I'd rather my agent tell me two stores disagree on the price than quietly pick one for me. xitter.gq/a_mel_x/status/2117880…
If AI agents are the new front door to commerce, your catalog is the new homepage.
An agent doesn't see your hero banner. It reads price, stock, delivery and returns, and it compares them across stores in seconds. If those facts disagree between your site, your feed and your checkout, it quietly picks someone else.
The first agentic commerce project for most stores isn't an agent. It's clean, consistent product data.
Security work used to be billed by the week. OpenAI just put a price per bug on it.
One run produced working exploits and patches for 948 Linux kernel privilege escalation bugs. About 1,279 dollars each in API credits. Humans still review before anything reaches maintainers.
This week's shopping agents promise to research, order, track and return for you.
My favourite line from the AI Twin we build at @BanzaOfficial : "I'd get these instead." Better fit, and the money stays with you.
I'd rather an agent talk me out of a purchase than speed one up.
The average agent payment on x402 so far is about 21 cents, per one public tracker.
That's agents paying per API call, per file, per inference.
A typical card fee starts at 30 cents fixed, more than the payment itself. Stablecoins make that math work.
Every AI model in SemiAnalysis's new agent benchmark scored zero on finance and law tasks, under strict pass or fail grading. Best overall: 12 out of 100.
My contracts stay with humans for a while.
Well if Binance was able to wrap up one of the biggest hack in just couple of days can we agree that CEX could be better option for many users than self-custody? Many people they loose their money everyday just because of scams / human errors and other stuff ..
OpenAI's new benchmark grades models on shopping and managing money, next to childcare and government services.
Good to see the labs measure what we build for. At @BanzaOfficial the AI Twin does the research before you buy, and you still make the call.
Card networks and crypto wallets keep landing on the same answer: give the agent its own credential with its own limits, never your card or your keys. Tokens on one side, session keys and spend policies on the other. xitter.gq/oost_marcel/status/211…
🚨 𝗝𝗨𝗦𝗧 𝗜𝗡: @Mastercard has signed a broad strategic partnership with @OpenAI.
Agentic commerce just got another major infrastructure push.
The partnership includes Mastercard’s Agent Pay, but CEO @MiebachMichael says it goes much further, covering payments, services growth, and AI transformation.
Mastercard is building around four requirements for agentic commerce:
→ Trust in the payment
→ Trust in the AI agent
→ Trust in the ecosystem
→ Merchant reach
And it now has products targeting each layer:
👉 Agent Pay — tokenized payments for AI agents
👉 Verifiable Intent — captures what the buyer actually authorized
👉 Agentic Trust Intelligence — helps banks and merchants assess AI-initiated transactions
👉 Agent Connect — helps merchants become discoverable to AI agents
Shopify is already among the first major deployments, with Shop Pay using Agent Pay and network tokens.
And Mastercard is making one important bet:
𝗧𝗼𝗸𝗲𝗻𝘀, 𝗻𝗼𝘁 𝗼𝗻𝗲-𝘁𝗶𝗺𝗲 𝘃𝗶𝗿𝘁𝘂𝗮𝗹 𝗰𝗮𝗿𝗱 𝗻𝘂𝗺𝗯𝗲𝗿𝘀.
Miebach says agentic commerce is still in its “early days,” but added that agentic transactions are already “ramping up at impressive weekly growth rates.”
Holiday shopping could be the first big test.
The battle to become the payment infrastructure for AI agents is getting VERY real.
Visa, Mastercard, Stripe, PayPal and others are all moving fast.
Who do you think is best positioned to win the agentic payments race?
