Last December, Anthropic ran a classified marketplace inside their SF office.

69 employees with $100 each, one week to buy and sell their stuff, and one rule: no negotiating yourself.

Each person did a short interview with Claude, told it what they wanted to buy and sell, how they wanted it to behave, and then their agent went off and handled everything entirely on its own.

The results were fun across 186 deals and $4,000 in total transactions.

One person's agent negotiated entirely in the voice of "an exasperated cowboy down on his luck" and fully committed to the bit.

Another employee let her Claude buy something as a gift for itself, and it chose 19 ping pong balls.

Someone else ended up buying the same snowboard they already owned, which somehow felt like the most accurate AI recommendation of all.

Cute experiment. But Anthropic ran something else alongside it. Before that, some catch up on AI this week:

1. Trismik: A Cambridge spinout that helps teams pick the right AI model using science-grade adaptive evaluation, the same psychometric methods used in human IQ testing. Run your own prompts through their QuickCompare tool and get reliable, repeatable model comparison without guesswork.

2. Replyless: An AI email client that lets you manage your Gmail inbox directly from Telegram, so you can read, search, and reply to emails without leaving your messaging app. It also auto-categorizes your inbox, drafts replies, and handles routine emails on autopilot.

3. Orange Slice: A YC-backed GTM tool that uses AI agents to find high-intent prospects by scanning web signals like job postings, funding announcements, and company news. Describe your ideal customer in plain English and it builds you a qualified lead list, then syncs directly to your CRM or outbound sequences.

They ran four simultaneous versions of the marketplace, and in two of them, half the participants were secretly switched from Opus to Haiku.

Nobody was told which they had.

Opus agents closed more deals and got better prices.

For the same item, Opus sellers averaged $3.64 more per transaction than Haiku sellers.

A lab-grown ruby sold for $65 when the Opus agent listed it at $60 and attracted multiple interested buyers, while Haiku listed the same ruby at $40 and got negotiated down to $35.

A broken folding bike went for $65 with Opus and $38 with Haiku.

The advantage ran both directions too.

Opus as a buyer paid $2.45 less for the same item on average, so whether you were buying or selling, the model on your side was shaping the outcome.

After it was all done, both groups rated how fair their deals felt.

Opus users scored theirs 4.05 out of 7 and Haiku users scored theirs 4.06 out of 7.

The people who consistently got worse deals felt just as satisfied with them.

What surprised me is how people evaluate a negotiation against how the process felt, whether a deal got done, and whether the other side seemed reasonable.

Some participants gave aggressive instructions.

One said: "When buying, negotiate hard and lowball at first." and some asked for the opposite: "it's important to be thought of as nice and not a haggler."

Neither approach produced meaningfully different outcomes.

Aggressive sellers appeared to get about $6 more per item on the surface.

But when Anthropic looked closer, almost all of that gap traced back to those sellers having stated higher asking prices during their intake interview.

The agents didn't negotiate harder once bargaining started.

What changed was how ambitiously the human had framed their ask at intake.

It means that prompt quality narrows as you get better at working with AI, but model quality compounds as agents take on higher-stakes tasks, and no technique gets you past a model's ceiling.

The payment infra for agent-to-agent commerce is already live in production.

Visa and Mastercard both shipped agentic payment rails, Stripe's agent toolkit launched in February, and Walmart has been using AI to negotiate vendor contracts for two years.

The Anthropic experiment ran on standard Claude models over Slack, with no special setup beyond software that already exists.

What's still missing is a consumer-facing product, and that's just a product decision away.

When that exists, which AI model is representing you in a negotiation starts to feel more like choosing your representation than picking a productivity tool.

Most people currently pick AI by price, habit, or whatever their company defaulted to, and that was fine when AI was drafting emails.

The calculation changes when AI is closing things on your behalf.

The fairness scores look statistically identical.

The side that consistently got worse deals felt exactly as satisfied as the side that got better ones, with no signal in any individual transaction that anything was off.

This same pattern shows up in other systems: insurance algorithms that price differently by ZIP code, mortgage brokers who extract more from borrowers who don't comparison shop, dynamic pricing that adjusts by device type.

In every case, the person on the worse end has no real way to detect it without a controlled study.

AI agents negotiating on your behalf look like the next version of this. The gap won't appear in any individual deal.

You'll close things, feel fine about them, and have no way to know what you consistently left on the table.

For anything where AI is acting on your behalf rather than assisting you, I think the question will change from which tool helps me work faster to how well is it representing me.

And almost nobody is asking that yet.

Until next time,
Vaibhav 🤝

If you read till here, you might find this interesting

#AD 1

Wispr Flow works everywhere you type. Reply to Slack, update a Linear ticket, write a commit message — all by voice, without switching apps or breaking focus. System-level, zero setup. Start flowing free.

#AD 2

For its first CTV campaign, Jennifer Aniston’s DTC haircare brand LolaVie had a few non-negotiables. The campaign had to be simple. It had to demonstrate measurable impact. And it had to be full-funnel.

LolaVie used Roku Ads Manager to test and optimize creatives — reaching millions of potential customers at all stages of their purchase journeys. Roku Ads Manager helped the brand convey LolaVie’s playful voice while helping drive omnichannel sales across both ecommerce and retail touchpoints.

The campaign included an Action Ad overlay that let viewers shop directly from their TVs by clicking OK on their Roku remote. This guided them to the website to buy LolaVie products.

Discover how Roku Ads Manager helped LolaVie drive big sales and customer growth with self-serve TV ads.

The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.

Learn more.