The deal you didn't know you lost

🔭 Everyone's worried about the wrong side of the agent gap

The consensus take after Project Deal is straightforward: stronger consumer agents will out-bargain weaker ones, the rich get richer, regulate fast. That's the take in most of the coverage. It's also probably wrong — not because the gap isn't real, but because it isn't durable on the consumer side.

Here's the chain. First, consumer-side agent quality is already compressing. Open-source models are two to three benchmark points behind frontier and shrinking. A well-prompted Llama running on Browser Use closes most of a one-shot negotiation gap. The “Opus vs Haiku” spread Anthropic measured was within one model family; cross-family on the same task, the spread is smaller and falling.

Second, the merchant side is going the opposite direction. A consumer agent sees one negotiation. Amazon's pricing agent — or any platform's — sees a billion. Every transaction trains the counter-agent. That's not a model-quality advantage that commodifies; it's a data advantage that compounds. The consumer agent goes to school once. The merchant agent goes to school every Tuesday. Saperly's agent-only phone rails and Salesforce's Headless 360 are the early infrastructure for this side of the table.

Third, the asymmetry shows up in pricing. Once a merchant's agent can identify a weaker counter-agent — by behaviour, by API fingerprint, by anchor strategy — it can quote a higher price for the same SKU. Not maliciously. Just optimally. That's already how dynamic pricing works for human signals like browser type and zip code. Agents make it cleaner.

Fourth, this becomes detectable. Not by the consumer — by regulators and class-action lawyers, who can run statistical tests across millions of transactions in a way no individual ever could.

So the prediction: within roughly eighteen months — call it Q3 2027 — the first major lawsuit lands against a platform whose merchant-side AI was found to systematically discriminate between consumer agents. The story won't be “AI made shopping better.” It'll be “your AI got profiled.”

Read the deep dive: The counter-agent is the moat →

🎯 PROMPT OF THE WEEK

The Agent Brief

The Project Deal interviews are the most interesting part of the experiment. Each participant got a ten-minute interview, and that interview built the system prompt that ran their agent. If a future agent is going to negotiate on your behalf — for purchases, salary, refunds, anything — your “agent brief” is the leverage you have.

Here's a starting brief you can save and reuse. Paste it into Claude or ChatGPT before any negotiation task:

You are negotiating on my behalf. Before I give you the

task, here's my profile — refer back to it whenever you

make a decision.

WHAT I VALUE:

- Outcome > speed. I'd rather wait two days for a 15%

better deal.

- Honesty > likeability. Don't soften my position to

be polite.

- Walk-away is real. If terms cross my floor, exit

cleanly.

MY DEFAULTS:

- Floor (worst acceptable): [fill in]

- Target (what I actually want): [fill in]

- Anchor (where to open): [fill in]

MY STYLE:

- Direct, not aggressive. No hard-sell language.

- British English. No exclamation marks.

- One-line summary at every decision point so I can

intervene.

WHEN UNCERTAIN: Ask me. Don't guess on price.

🛠️ TOOL OF THE WEEK

Browserbase

The browser stopped being for humans. ★★★★½ / 5

Skip if: you don't write code or commission anyone who does. Use if: your team is building anything that needs to log into sites, fill forms, or scrape behind auth walls — internal tools, monitoring, lead enrichment, anywhere Zapier hits a paywall and a human has been doing the click-through.

Browserbase is the cloud-browser layer most agentic startups now build on. Drop-in Playwright-compatible. They processed 36.9 million unique browser sessions in March, closed a $40M Series B at $300M last summer.

Describe it to a colleague: “It's a browser as a service — except the user is your agent, not you.”

Persistent sessions with managed cookies — no logging in twenty times.

Built-in stealth mode for sites that block bots reflexively.

Session recordings — watch what your agent did when something goes wrong.

Native integrations with Stagehand, Browser Use, Vercel's agent-browser CLI.

Free tier to start.

Try Browserbase →

💡 TIP OF THE WEEK

Why model choice beats prompt-tuning

Embarrassing admission: I have a Notion folder with 40+ saved prompts I've spent hours fine-tuning. It's possible most of that time was wasted. Project Deal just put a real dent in the prompt-engineering religion.

Anthropic ran identical interviews through Opus 4.5 and Haiku 4.5. Same instructions. Same data. On their negotiation task, Opus runs averaged $2.68 more per item for sellers and closed two more deals on average (Anthropic's published figures). The instructions barely mattered.

That's one task on one experiment — but the pattern holds anywhere the work involves ambiguity, multi-turn reasoning, or judgement under incomplete information. For those task types, model tier is doing more work than prompt-tuning. Cost difference between Haiku and Opus: roughly 5x per token. Usually worth it.

When it doesn't apply: narrow tasks — extraction, classification, summarisation. Smaller models match larger ones once the task is well-defined. Prompt-tuning still pays.

Pro move: before committing a recurring task to a model, run ten samples through both tiers on outcomes you actually care about — closed deals, accuracy on edge cases, customer satisfaction. Not benchmark scores. If the bigger model doesn't win on what matters, downgrade.

P.S. The deep dive on the counter-agent moat goes further on the merchant-side argument.

Read it →