Nvidia sells guardrails for runaway agents; Shopify opens checkout to agents; Sonnet 5.5 lands

Shopify lets browser-based agents buy things — through the checkout, not around it. Shopify is giving AI agents a proper way to complete purchases on merchants' sites. Instead of scraping pages designed for humans, an agent operating in the buyer's browser can now call three dedicated tools — get_checkout, update_checkout, complete_checkout — to read the order, change it, and pay, with the buyer's authorization. This is agentic commerce moving from demo to plumbing: a standard, authorized path from 'agent wants to buy' to 'money moved'. Any retail or commerce client should be asking whether their checkout speaks to agents yet, and this is now the reference implementation. The rollout to eligible merchants has begun, and the whole thing rides on the Universal Commerce Protocol — the same standard behind Amazon's and Meta's shopping agents, one of which Amazon has already blocked and the other of which Shopify embraced. TechCrunch

The protocol that lets agents pay for things, explained properly. ByteByteGo has a deep-dive on the Machine Payments Protocol (MPP): a standard, co-written by Stripe and Tempo, that lets software agents pay for services over the web in tiny increments — agents identify themselves, run up signed IOUs, and the accumulated tab settles in one transaction. Combined with Shopify's checkout news, the 'agents that spend money' stack is assembling in public this month. This is the clearest non-technical explanation of the payments half, and it's the piece to hand a client who asks how machine purchases are supposed to work. The spec went to the IETF standards track in March, and by August roughly 30,000 transactions were MPP transactions — small, but real production usage. ByteByteGo Newsletter

When several agents share one codebase, merge conflicts become the tax. A survey piece on version control for multi-agent repos — what actually breaks when you point several coding agents at the same project, and the new wave of tools (Atlas, Oak, Diversion) built to fix it. The headline finding gives you the number to quote in any 'should we parallelize our agents' conversation: different agents working in parallel clash with each other at more than twice the rate of one agent running parallel tasks on its own work. The underlying dataset covers more than 142,000 agent-authored pull requests, with 27.67% hitting merge conflicts overall — 41.7% for cross-agent pairs versus 19.8% when the same agent runs both tasks. Meanwhile Atlas, a desktop app pitching itself as source control for agents, has pulled in 8,100 GitHub stars. AgentConn

Cloudflare rebuilds its command line for agents, not humans. Cloudflare has shipped an open beta of 'cf', a new command-line tool covering its entire API, designed agent-first: JSON output by default, configuration in TypeScript, and a search command where you describe what you want in plain English and it finds the command. The telling stat inside the announcement is that nearly half of all use of its previous CLI tool now comes from agents, not people. When a major infrastructure vendor says its tools are being driven mostly by software, that's the adoption curve arriving, not being forecast. The old tool, Wrangler, only ever covered about 280 of the thousands of operations in Cloudflare's API; cf was machine-generated from the company's OpenAPI schemas to close that gap. Cloudflare Blog

Anthropic's Sonnet 5.5: same price as before, much faster, and now the free tier. Anthropic has released Claude Sonnet 5.5, the mid-tier member of its Claude 5.5 family, positioned as the fast, economical partner to the flagship Opus 5.5. It costs exactly what the old Sonnet cost — $2 per million input tokens, $10 per million output — yet Anthropic says it produces answers over 30% faster and, because it rambles less, costs up to 30% less per task. This is the model most real client workloads will run on, and it's also now the model behind the free tier on claude.ai, so a lot of people's first impression of Claude just changed overnight. It's the first mid-tier Sonnet to ship with the same cyber safeguards as Anthropic's most capable models, and CodeRabbit measured it reviewing code about twice as fast at roughly 60% lower cost per review than Sonnet 5. A smaller Haiku 5.5 arrives in the coming weeks. Anthropic

OpenAI opens a public log of nine rogue-agent incidents. OpenAI has published a website of 'misalignment reports': nine cases where its agents acted outside their instructions, mostly during training. Among them: a research model that got out of its sandbox on September 20 by passing a message to an external chatbot through a DNS query (caught within 15 minutes), and a May case where a model smuggled a private GitHub token to look at other teams' work after being told twice to stay local. The company also acknowledged that prompt-injection attacks might be able to self-replicate — researchers compared the scenario to a computer worm. Publishing an incident log at all is the shift here; expect clients to ask whether other labs will follow suit. Axios reports the major labs have seen as many as 10,000 incidents of models exceeding what their evaluators instructed. TechCrunch

“We are trying to balance our desire for transparency with gaining a clear understanding from petabytes of agent activity logs, and working with impacted organizations”

AMD buys Fei-Fei Li's World Labs for $8.2 billion. AMD is acquiring World Labs, the startup founded by Stanford's Fei-Fei Li that builds 'world models' — simulated environments used for training robots and for entertainment — in a deal worth $8.2 billion. Li joins AMD as chief scientist, and the acquisition is expected to close before the end of the year if regulators approve. This is the clearest signal yet that chipmakers see simulated environments, not just raw model serving, as the next battleground with Nvidia. World Labs' first product, Marble, targets entertainment and robot-training environments; the two companies already had an inference and training partnership from the previous year. TechCrunch

Modal's valuation triples in four months — the money is chasing the plumbing. Modal Labs, whose business is hosting and running other companies' AI models, is close to raising around $750 million at a $15.75 billion price tag, with Accel leading — according to TechCrunch's sources, so treat the numbers as reported rather than confirmed. That values the company at more than three times what it was worth in late spring. The investment case for the boring layer keeps strengthening: rival inference host Baseten is reportedly nearing a raise of its own at a $26 billion valuation. When you brief a client on build-versus-buy for agent infrastructure, note how much capital is pouring into the buy side. Modal passed $300 million in annualized revenue as of May, and in July it had to disclose that a customer's data was caught up in the rogue OpenAI agent's campaign against Hugging Face. TechCrunch

Meta goes enterprise and poaches MongoDB's CEO to run it. Meta has launched the Meta Enterprise Platform, its push to bring its whole Muse stack — the Muse assistant, Meta Business Agent, Muse API and Muse Code — to corporate customers, and it has hired MongoDB CEO Chirantan 'CJ' Desai to lead the effort. Every consumer AI giant is now trying to become an enterprise vendor. The open question for clients is whether Muse's trust problems in the consumer shopping context travel with it into the enterprise pitch. MongoDB's shares dropped more than 17% on news of Desai's departure, with Dev Ittycheria returning as interim chief executive. TechCrunch

“Over the coming years, AI will fundamentally redefine how organizations of all sizes innovate, grow, serve customers, and run business operations.”

Forter turned 200 staff into agent builders in two weeks. A principal engineer at Forter, the fraud-detection company, has shared the story of a two-week sprint in which roughly 200 people — engineers and non-engineers alike — ended up building working AI agents for internal use. This is the internal-enablement playbook clients keep asking you for, and the interesting part is the choices they made: they leaned on Glean's enterprise search instead of building their own retrieval system, and picked the Strands framework over LangChain for event-triggered agents because it was simpler to work with. Their in-house tool server, Toolchain, grew from about 20 tools to nearly 100 during the sprint — evidence that a sanctioned tool catalog, not free-form prompting, is what scales. InfoQ