Meta, Walmart and Stripe publish the Personal Agent Protocol
The agent economy is writing its rulebook tonight: one open standard for AI bots at the checkout, a cheaper image model from Google, and Anthropic turning bug-hunting into a tiered product.
Meta, Walmart, Stripe and Sierra are publishing an open "personal agent protocol" — a standard that defines how personal AI agents interact with businesses online. The group behind it reads like a cross-section of agentic commerce: Meta, Sierra, Walmart, Stripe, Shopify, Genesys, Rocket, NiCE, Decagon and Instinct, with Sierra co-founder and OpenAI chairman Bret Taylor leading the effort. The protocol handles the parts that are currently improvised — authenticating that a visitor is an agent acting for a real person, letting the business grant read-only or write access, and giving companies visibility into what the agent actually did — built on OAuth, with the first spec version due later this month and open for anyone to implement. The timing is the story: a month after Meta's Muse went viral, Amazon has blocked it, airlines and marketplaces keep agents out, and even Walmart's own bot defenses have tripped over agents that were technically welcome. Taylor's line sums it up — "It is kind of chaos until such a standard exists" — and Meta frames the stakes commercially: turning away a personal agent now means turning away the customer behind it. Our take: standards rarely win by being correct; they win when the biggest traffic sources adopt them, and OpenAI and Anthropic are still missing from the roster.
Google cut the price of its workhorse image model roughly in half with Nano Banana 2.1. The new model replaces Nano Banana 2 from February, runs on Gemini 3.6 Flash, and drops the cost of a 1K image from 6.70 cents to 3.36 cents (4K images fall from 15.10 cents to 7.56 cents), with batch pricing halving it again. Google's own evaluations show big gains — infographic accuracy up from 0.179 to 0.521, and multi-character consistency well ahead of both Nano Banana 2 and the pricier Pro — but those are Google's side-by-side human ratings, not an independent leaderboard, and The Decoder's hands-on test still found Nano Banana Pro produces noticeably better images. The practical read: 2.1 is the cheap, fast default now, Pro remains the quality tier, and the previous model shuts down on October 29.
Anthropic merged its cyber programs into one expanded, tiered product — and opened its newest models to security teams. The company folded Project Glasswing into a bigger Cyber Verification Program with three access tiers, including access to its latest models; separately, Anthropic says its own hunting turned up more than 5,500 verified vulnerabilities between April and October, with Glasswing partners finding over 129,000 between April and July, more than 33,000 of them critical or high severity. Bloomberg and The Information both covered the access expansion. Coming days after the Pentagon said it had stopped using Anthropic products, the direction of travel is clear: frontier-model access for cyber work is becoming a line item Anthropic sells directly, to buyers who choose it.
What to watch: whether OpenAI and Anthropic sign on to the protocol before the spec hardens — and who moves first.
Is an open standard enough to stop the bot walls from going up everywhere, or will the biggest sites keep blocking agents anyway? Tell us in the comments.
