The Collection

Crane’s DealBook letter names the shops helping brands show up when AI shopping agents decide. Limy, founded by Aviv Shamny and backed by a16z. Gauge, founded by Caelean Barnes. New Generation, founded by Jonathan Arena. The product is not another banner. It is readiness for a buyer that does not feel dopamine.

Shamny’s printed contrast carries the claim: advertise to humans and you buy eyeballs, emotions, dopamine; advertise to agents and you buy mathematical logic. Adobe’s printed figure: AI-source traffic to US retail sites up 393% in 2026. Morgan Stanley’s printed forecast: by 2030, agent-influenced spend might be up to 20% of US e-commerce, or $385 billion. Keep those as printed, not proven.

A chat buy is not a brand visit.

RTA Store CEO Tyler Ackerman paid $15,000 to New Generation for AI visibility after the brand stopped showing in LLM recommendations. He envisions up to $12,000 a month later. New Generation and Visa rate brand sites from 1 to 100 for agent readiness. Jeff Blackman of the Barbarian Group prints the end state: more and more you will just click buy in the AI chat and never go to the brand site. Toby Coulthard of Jacquard prints the brake: models change faster than the ability to game them; the work returns to fundamental marketing. Amazon is blocking some AI agent shoppers. Many retailers still want the AI traffic. Sunday files the pitch without inventing conversion rates the letter did not print.

Krazit’s Runtime letter puts Cloudflare Basin generally available: end-to-end analytics on the Developer Platform, Apache Iceberg, pipelines, catalog, R2 SQL, and free egress for portability. The stack claim is not a new warehouse brand. It is analytics that stay beside the compute and the object store, with egress that does not trap the data.

The same sheet carries MongoDB 9.0 and Atlas Infinite, with Dev Ittycheria returned after CJ Desai left for Meta. AWS released Strands Decider 2B under Apache 2.0: roughly 2 billion parameters, answers in tens of milliseconds, fine-tuned from Qwen. Runtime frames it as jumping on the decision-model bandwagon after TypeSafe Jev. Sunday keeps that framing as printed and does not invent latency tables the letter did not carry.

Free egress is a portability claim.

Quote of the week, as Runtime prints Forrester’s Jeff Pollard on harnesses: people still do not know what they are. The Collection files Basin beside The Pitch. A shopping agent needs a stack that can answer in tens of milliseconds. A brand that pays for agent readiness still needs someone who can say what a harness is for.

Peter Norvig, in a Turing Post interview dated 3 October, puts the limit in one line: “I think trust can only last so long… corrupting power… concentration of wealth and power.” The optimism he keeps is on open models. DeepSeek, Kimi, and the rest, in his telling, mean power more widely distributed, and therefore safer.

On agents he is plain about the shift. His book was “the intelligent agent book.” Now the main interaction is swarms, not one agent. Safety, as he frames it, is figuring out the right level of control and software engineering practices, with or without recursive self-improvement.

Trust can only last so long.

Sunday does not invent Norvig’s full transcript past the verified lines. It files the trust claim beside The Pitch: a brand that sells to a bot still needs a public that trusts the swarm enough to click buy. Concentration of wealth and power is the corrupting twin of a chat that never opens the brand site.

Kilo Code and Anaconda, via Substack around 2 October, print a new product equation: (Cost + Performance + Efficiency) × Security. The multiply is the point. A model that is cheap, fast, and efficient can still score near zero if the security term collapses.

The same week’s safety sheet, as printed: Google phased the Gemini 4 Argon rollout for safety (WSJ). OpenAI shelved Astra and shipped GPT-6.1 Sol at DevDay, citing safety. Anthropic’s Claude Sonnet 5.5 carries stronger safeguards; higher-risk cyber requests are refused. Sunday does not recycle Saturday’s Anthropic IPO, leap-of-faith, or non-cancellable invoice stories. Those stay on The Leap.

Security is the multiplier, not a line item.

Parallel’s September updates sit as a short shelf notice: Data Connectors for trusted sources; a Search Capability Leaderboard; Parallel Search inside LangChain Managed Deep Agents; a Lovable connector. Kept as named shipments, not invented metrics.

LawZero / Scientist AI, filed via Medium by Roan Brasil Monteiro, is MEMBER-ONLY. Unread paywall. The open snippet only: fifteen months after Bengio’s $30 million LawZero launch; an about 40-page paper; up to CAD $300 million from Canada and Germany; a system that estimates P(harmful action) and blocks above a threshold. Sunday does not invent the rest of the paper from a teaser.