The deal you didn't know you lost

In Today’s 5-Minute AI Digest. You will get:

1. The MOST important AI News & research
2. AI Prompt of the week
3. AI Tool of the week
4. AI Tip of the week

…all in a FREE Weekly newsletter. 

THE AI NEWSLETTER THAT TURNS NEWS INTO ACTION
ISSUE #135 WEDNESDAY · 6 MAY 2026

THE HOOK

Two Anthropic employees sold the same broken folding bike to the same buyer. One pocketed $65. The other got $38. Same item, same buyer — only the AI agent doing the haggling changed. Neither side noticed the gap. If your business is heading toward agents handling renewals, vendor pricing, or sales outreach — and most are — this is your problem now. Both sides rated the deals fair. That's the part that should keep you up.

AI NEWS ROUNDUP

This week in AI

1 Anthropic published Project Deal. Sixty-nine employees, 500+ items, 186 deals, $4,000 changing hands inside Slack — all negotiated by Claude agents with no human input after the intake interview. The published writeup details four parallel runs comparing Opus 4.5 and Haiku 4.5 agents on identical preferences. Worth the ten minutes if you haven't read it. Read the writeup →

2 The Trump White House is reversing course on AI oversight. Sixteen months after rescinding Biden's safety-testing executive order, the administration is now drafting a new one — a working group that would review frontier models before public release. The trigger? An unreleased Anthropic model called Mythos that the company itself flagged as potentially capable of triggering “a cybersecurity reckoning.” Officials reportedly briefed Anthropic, Google and OpenAI last week. NYT via Bloomberg →

3 Time-to-exploit has gone negative. Mandiant's M-Trends 2026 report (published April, sample of 600+ investigations) found 28.3% of disclosed CVEs now exploited within 24 hours of disclosure. The “negative” part is a small but growing category: exploits landing before the patch ships. Caveat — this isn't most CVEs, just the most weaponisable ones. But Black Hat Asia put a separate number on the trend: bug-to-working-exploit collapsed from five months in 2023 to ten hours in 2026, with frontier models doing the offensive work. The Hacker News →

4 Saperly launched as the first phone carrier built for AI agents. Real phone numbers, voice, SMS, routing — all through one API designed for agents instead of humans. Your agent gets its own line, owns its own caller ID, switches between calling and messaging without a second provider. Boring infrastructure. The kind that becomes load-bearing about six months before anyone notices. Saperly →

OUR ANGLE

THE THREE SPECIALS

Do · Use · Understand

Why it works: Project Deal showed that clarity of preference beat tactical instructions. Telling the model your floor, target and anchor — three numbers — does more than a paragraph of “be tough but friendly.” Anchoring is also the most under-used negotiation lever; if you don't tell the agent where to open, it'll start somewhere reasonable, and reasonable usually loses.

Where to be careful: this brief is built for low-to-medium stakes, mostly-text exchanges — refunds, second-hand purchases, vendor renewals, freelance scope. It is not suitable as-is for salary negotiations, legal contracts, or anything where the counterparty is human and the relationship matters more than the price. For those, use it as a thinking aid, not an execution tool.

Works best on: Claude Opus 4.5, GPT-5.

YOUR MOVE

Pick one. Reply by Friday.

You just learned:

Project Deal showed agent quality moves real money — and losers don't notice.

The bigger moat probably isn't your agent. It's the merchant's.

Model tier does more work than prompt-tuning on tasks with real stakes.

Pick one of these three and do it before Friday. Save the Agent Brief, run the Opus-vs-Haiku A/B on a recurring task, or read Anthropic's Project Deal writeup with our angle in mind.

Then reply and tell me which one. That's the action that matters this week — not the share, not the deep dive. Reply. I read every response. The pattern across replies is what shapes next week's issue, and the readers who actually move tend to be the ones whose names I start to recognise.

—

R. Lauritsen