Anthropic ships Claude dreaming, AWS Bedrock AgentCore launches stablecoin payments with Coinbase and Stripe, CAISI tests Google, Microsoft and xAI models, TikTok Smart+ adds module controls

Anthropic used its Code with Claude conference in San Francisco to ship dreaming, multiagent orchestration and outcomes for Claude Managed Agents. Dreaming is a research preview; the other two are public beta. Harvey says completion rates rose roughly six-fold with dreaming on. The interesting part for operators is not the science but the architecture — agents that compound learning between sessions start to look more like staff than tools.

AWS launched Amazon Bedrock AgentCore Payments in preview with Coinbase and Stripe. Agents now hold a wallet — Coinbase or Stripe Privy — and pay for APIs, MCP servers, paywalled content and other agents in USDC, settling in roughly 200 milliseconds on Base or Solana. Thomson Reuters, Warner Bros. Discovery, Cox Automotive and PGA Tour are testing. This is a foundation move, not yet a consumer checkout product.

The Center for AI Standards and Innovation (CAISI) signed pre-deployment testing agreements with Google DeepMind, Microsoft and xAI. The OpenAI and Anthropic memoranda were re-papered. Bloomberg, Washington Post and Tom's Hardware all report the White House is drafting an executive order to formalise model vetting, with Anthropic's Mythos cyber-capable model named as the catalyst. The tone has shifted from deregulation to FDA-style review.

TikTok published its Q2 2026 product preview. Smart+ is no longer all-or-nothing — advertisers can switch automation on or off at module level for targeting, budget and placements. Pulse Mentions and Tastemakers expand to Canada and a wider US pool. Collage Carousel puts a hero plus three product images in the first frame, each clickable to a PDP. WhatsApp and Messenger become eligible follow-up surfaces.

Watchlist: Gemini 3.1 Flash-Lite went generally available on 7 May at $0.25 per million input and $1.50 per million output — priced as a serious workhorse for content moderation, classification and bulk generation. And Stripe Sessions left Klarna BNPL inside Shared Payment Tokens, giving agentic checkouts their first real BNPL rail.

The pattern across the day is concrete. Agents that learn, agents that pay, agents that get audited by government, and a paid-social platform finally giving operators a wrench rather than a black box.

Transcript

SPEAKER_00 0:01

Welcome to the Daily Marketing Brief. Your daily AI news and tactics for marketers who move fast. I'm your host, Jen Bryan, and here's today's update. The defining tension this week is the gap between what AI agents can now technically do and what operators can actually trust them to do without supervision. That gap narrowed on three different fronts in the last 48 hours. Better long-running memory, real money and agent wallets, and a government that wants a look at frontier models before you ever see them. The fourth story is much more practical. TikTok finally letting paid teams turn off the bits of smart plus they have been complaining about for a year. In today we cover Anthropics code with clawed announcements and what dreaming actually changes for agency and in-house teams using AgenTic workflows, the AWS Coinbase and Stripe Agent Core Payments launch, and why this is the plumbing layer for Agentich Commerce, not the storefront. CAISI's pre-deployment testing agreements with Google, Microsoft and XAI and the executive order being drafted in the background, and TikTok's Q22026 product preview, which is the one with the most immediate effect on how performance teams run accounts on Monday morning. First up, Anthropic ships dreaming outcomes and multi-agent orchestration for cloud managed agents. What happened? At the Code with Cloud Developer Conference in San Francisco, Anthropic announced four updates to Cloud Managed Agents. Dreaming, Outcomes, Multi-agent Orchestration, and Webhooks. Dreaming is a scheduled background process that reviews an agent's prior sessions, extracts patterns, and curates its memory store between runs. Outcomes lets you define what good looks like for a task and have agents optimized to it. Multi-agent orchestration lets up to 20 parallel specialist agents coordinate on a problem. What is confirmed? Anthropic published the announcement and VentureBeat, SiliconANGLE, the new stack, and 9 to 5 Mac all carry primary detail. Dreaming is in research preview. Outcomes, multi-agent orchestration and memory are in public beta. Anthropic states that dreaming does not modify model weights, it works only on agent memory stores. Harvey, the legal AI vendor, is on record reporting that task completion rates rose roughly sixfold once dreaming was enabled in tests, particularly for file type workarounds and tool-specific patterns. Why it matters? The constraint on serious agent deployments has been continuity. Agents that forget what they learned yesterday cannot replace a junior analyst, a paid media coordinator, or a paralegal. Dreaming is the first credible attempt to let an agent improve overnight, the way a person rereads their notes on the train. What it means for business. If you are running pilots on Cloud or another agent platform, the operating model shifts from prompt engineering to memory hygiene. Who curates the memory? What gets surfaced to the team wide store? What gets purged? You now have an agent ops problem, not a prompt writing problem. What it means for marketers and agencies, three places this matters. One, briefing repositories, your agency tone of voice, channel rules, and previous campaign learnings become the corpus an agent dreams against. Two, paid media, long-running campaign agents with persistent memory could finally hold context across reporting cycles instead of losing the thread every Monday. Three, content workflows, sub agent crews of researcher, drafter, editor, and fact checker working in parallel become a viable pattern, not a demo. My take. Sixfold completion rate is one customer in controlled conditions. Treat it as directional, not predictive. The bigger signal is that Anthropic is now building for fleets of agents, not single chatbots. If your AI roadmap still says deploy a chatbot in Q3, you are a strategy generation behind. Confidence level, high on the announcement and feature set, medium on the Harvey numbers, single source vendor shared, low on how robust dreaming will prove outside research preview, especially with security-sensitive memory. Up next, AWS launches Amazon Bedrock Agent Core Payments with Coinbase and Stripe. What happened? On May 7th, AWS announced Amazon Bedrock Agent Core Payments and Preview built with Coinbase and Stripe. AI agents can now hold a wallet, either a Coinbase wallet or a Stripe Privy wallet, and autonomously pay for APS, data feeds, MCP servers, paywalled content, and other agents. Settlement is in USDC on base or Solana in roughly 200 milliseconds. What is confirmed? AWS published the announcement on its machine learning blog, Coindesk the Block, Finance Feeds, and Coindesk's Enterprise Feed carry the primary detail. The protocol layer is Coinbase's X402 and HTTP native standard for agent-to-agent stablecoin transactions. Named enterprise testers are Thomson Reuters, Warner Brothers Discovery, Cox Automotive, and PGA Tour. Initial scope is micropayments for APIs, paywalled content and machine-to-machine purchases. A WS says it intends to expand to larger transactions such as hotel bookings, travel reservations, and merchant payments in future versions. Why it matters. Up to now, agents could fetch but not buy. The Agentic Commerce conversation has been mostly about end-consumer checkouts inside ChatGPT or Gemini. Agent core payments is the other half of the system, agents paying their own way for the services they consume to do their job. Without that, every meaningful agent had to be wrapped in a human controlled credit card. What it means for business, two things. First, machine readable pricing is about to become a real product surface. If your data, content, API, or tool can be priced and consumed by an agent, you can sell it without going through a human procurement loop. Second, enterprise CFOs now have to think about agent spend categories, and your finance team is not ready for line items denominated in USDC. What it means for marketers and agencies. Direct application is limited today because consumer checkout is still next phase. But there are two adjacent implications. One, paid content gates. If your client is a publisher or a SAS, machine readable paywalls become a revenue line. Two, the Stripe Sessions partnership earlier in the month already lined up Klarna BNPL inside Stripe's shared payment tokens. So the consumer side of Agentic checkout is being assembled in parallel. Treat AgentCore Payments and Stripe SPTs as the same story. Agentic Commerce is being plumbed in layer by layer. My take stablecoin micropayments is not the headline. The headline is that agents have wallets and someone now decides what they can spend. That is a control plane problem. Your engineering and finance functions are about to inherit. Resist the urge to launch crypto pilots. Do learn what X402 is. Confidence level, high on the launch and partners, medium on adoption and preview status, enterprise testers only, no public usage data, low on near-term revenue impact for most marketing teams. Up next, CAISI signs pre-deployment testing agreements with Google, Microsoft, and XAI. White House drafts AI vetting executive order. What happened? The Center for AI Standards and Innovation, which sits inside the U.S. Department of Commerce, announced agreements with Google DeepMind, Microsoft and XAI to evaluate frontier AI models before public release. The existing OpenAI and Anthropic memoranda from 2024 have been repapered to reflect CAISI's directives under Commerce Secretary Howard Lutnick. Separately, Bloomberg, The Washington Post, and Tom's Hardware report that the White House is drafting an executive order to formalize pre-deployment vetting, with National Economic Council Director Kevin Hassett comparing it to FDA drug approval. What is confirmed? The CAISI agreements with Google, Microsoft, and XAI are publicly announced. CAISI says it has completed more than 40 evaluations to date, including on unreleased models. The executive order is reported by multiple outlets but has not been signed. A White House official told the New York Times any announcement would come from the president directly. The named catalyst is Anthropic's Mythos model, reported to materially upgrade the speed and sophistication of code level vulnerability discovery. Why it matters? This is a meaningful policy reversal. The current administration entered office revoking Biden era AI safety executive orders. Mythos has dragged it back toward pre-deployment review. The signal is that frontier model release timing is now a function of regulatory clearance, not just product readiness. What it means for business, if you procure AI from frontier vendors, expect launch slippage and tighter NDAs around model behavior. Enterprise buyers should ask vendors about CAISI evaluation status as a procurement question. For anyone building on top of Frontier APIs, factor regulatory drag into roadmap assumptions for the second half of 2026. What it means for marketers and agencies, indirect but real, model launch cadence drives the cadence of new ad products, search features, and creative tools. If Frontier launch is slow by a quarter, expect the AI feature pipeline at Google, Microsoft Meta, and TikTok to slow with it. The other angle is reputational. Brands deploying AI agents will increasingly be asked by procurement and by press whether the underlying model has been evaluated. My take. Confidence level. High on CAISI agreements. Medium on the executive order, which is drafted not signed. Medium on Mythos as catalyst and well sourced but partially based on briefings. Lastly, TikTok Q2 2026 product preview. Smart Plus Module Controls, Pulse Expansion, Collage, Carousel, What Happened? TikTok for Business published its Q2 2026 product preview. Three meaningful changes. One, Smart Plus is moving from all or nothing automation to module level control. Advertisers can switch automation on or off independently for targeting, budget, and placements with a Smart Plus label visible against each module. Two, the Pulse suite is being rounded out. Pulse Mentions extends to Canada in Q2. Pulse Tastemakers becomes available to a wider US advertiser base, joining Pulse Premiere and Pulse Core. 3. Collage Carousel, a new commerce ad unit showing one hero image and three additional product visuals in the first frame, each independently clickable to a product detail page. WhatsApp Messenger and other messaging apps become eligible follow-up surfaces from TikTok ads. What is confirmed? TikTok's official for business blog, PPC Land and SocialB carry consistent coverage. SmartPlus automatic placement now allows manual placement selection at campaign setup. The Q2 timing for mentions and tastemakers is explicit. Q3 timing was floated for further expansion. Why it matters? Smart Plus has been the Meta Advantage Plus analog and it has had the same complaint. Operators dislike opaque automation when budgets are non-trivial. Module level control is the first concession that performance teams have been asking for in earnest since launch. Collage Carousel acknowledges the obvious product carousels work, single product hero ads under monetize category-led shopping intent. What it means for business, for e-commerce in particular, the collage carousel format is worth a control test against single product video creative inside the same campaign. Smart Plus module controls let CRO and acquisition teams keep automation where it adds value and pull it out where they have stronger first party signals. What it means for marketers and agencies, practical. Three things. First, audit which Smart Plus modules you actually want automated for each client. Targeting may be worth keeping automated, placements may not. Second, brief creative teams on collage carousel, it changes the asset count and the briefing template. Third, the new messaging app handoffs open conversational retargeting flows that previously required clunky third-party stitching my take. This is the most operator-friendly TikTok product release in a year. Not glamorous, but it removes friction in real campaigns. The collage carousel will be most useful for retailers with broad ranges and weakest for single product brands. Confidence level. High on the announcement, medium on impact, performance lift will depend on category and creative quality, and TikTok has not published controlled test data. Two things to watch today. Gemini 3.1 flashlight went generally available on May 7th at 25 cents per million input and $1.50 per million output, with thinking levels exposed in AI Studio and Vertex AI. Google reports 2.5 times faster, time to first answer token versus the previous flash and 45% faster output. The point for operators, this is a serious workhorse tier model for content moderation, classification, large scale content drafting, and product copy variants. If your AI cost line is dominated by flash usage, this is worth a benchmark. Stripe sessions on April 29th locked Klarna BNPL into Stripe's shared payment tokens for US merchants already live with Klarna through Stripe. With Agent Core payments now arriving and Universal Commerce Protocol Partners expanding, the Agentic checkout layer is steadily becoming a commerce grade rail. Watch which BNPL options get added next. This is plumbing that will eventually decide whether agentic baskets convert at parity with the manual web. What matters most? Three of today's four stories, Anthropic, AWS, and CAISI, are about agents becoming serious infrastructure. Memory, money, and oversight are now shipping in parallel. The fourth, TikTok, is a reminder that the day-to-day commercial wins still come from operating leverage updates inside the platforms operators already use. The signal is that agentic capability is real and accelerating. The noise is the suggestion that any of this immediately replaces a marketing team. None of this replaces a marketing team. It changes what the team spends its hours on and what I would do this week if I were running this account, brand, or agency. Audit your smart plus campaigns this weekend. Identify modules where automation is hurting and switch them off as soon as the Q2 controls roll out. Document baseline C, PA, and ROAS now so you can measure the change. Brief creative for Collage Carousel. Source one strong hero plus three category spanning product visuals per priority SKU range. Aim to test in the first two weeks of availability. If you are running an AI agent pilot, write a one-page memory policy. What gets stored, who curates it, what gets purged, and what stays out of the team-wide store entirely. Do this before Anthropic Streaming gets out of research preview into your stack. Add a Gemini 3.1 flashlight benchmark to your existing AI cost stack. Pick one bulk task and product description rewrites, review classification, add copy variants, and run a paired test against current models on cost per acceptable output. Write a two-line procurement question for your AI vendors. Which models are CAISI evaluated? And what is the disclosure timeline if a frontier model is delayed for review? Send it to legal security and your AI vendor account managers. What I tell a client today, Agentic Commerce is being assembled. Wallets at AWS, BNPL at Stripe, check out at Google and OpenAI. Plan for it, do not build for it yet. TikTok's Smart Plus now gives you controls. If your account has been suffering since automation, this is the moment to reset. The most valuable AI work this quarter is curating the inputs your agents will dream against, not adding more agents. The story of the week is not a model launch. It is that AI agents have started to look like employees in three boring important ways. They remember they have a budget and they are starting to be vetted before they show up. Operators who set up the org around that idea this quarter will pull ahead of operators still treating agents as features. That's been today's episode. See you tomorrow.