Meta shipped a consumer agent that shops and pays for you, and within two weeks the largest store on the internet locked it out. That is the shape of September for anyone building on frontier models. This month's Roundup covers Amazon's block of Meta's Muse agent and the access question it forced, the model releases that cut frontier pricing again, the AWS runtime rebuild that makes always-on agents cheap to run, and the market moving its money toward agent identity. Each section connects the news to a decision a technology leader has to make this quarter.

Last month the story was a frontier model that advanced mathematics and tripped its own safety framework, plus an AWS control plane arriving to govern agents that enterprises were already deploying faster than they could watch. This month a personal agent walked up to the front door of the biggest storefront on the web and got turned away, while the same company showed what a governed agent is allowed to do instead. Here is what September delivered, and what it changes if you are putting agents into production.

eCommerce automation is here, but the Market needs to adapt

Meta launched Muse, a personal agent that shops, books appointments, and pays by signing in to sites as you and carrying your credentials along the way. It only took a couple weeks for Amazon to block it from the store. Meta never told Amazon the agent would shop there, Muse does not identify itself as an agent when it browses, and it appeared to capture and store customer credentials. Three days later Amazon shipped the opposite approach: the Selling Partner connection, which lets Claude and Amazon Quick act on seller accounts through access Amazon runs itself, with seller approval, scoped limits, and audit trails.

That is the question anyone building an ecommerce agent now has to answer. Does the agent reach a merchant through a declared, scoped, revocable connection, or does it log in as the user and scrape the page? An agent signing in as the person also leaves you no way to separate its actions from theirs in the log, and no way to shut it off without changing that person's password.

The infrastructure for the governed path is consolidating at three layers, and they are worth tracking as separate decisions:

• For discovery and checkout, the Agentic Commerce Protocol from Stripe and OpenAI and the Universal Commerce Protocol from Google and Shopify reach different surfaces and do different jobs.
• For payment, Mastercard Agent Pay and Visa's tokenized credentials bind a card to a specific agent, merchant scope, and spending limit, while AP2 handles signed mandates under the FIDO Alliance.
• For identity, per-agent credentials that an agent presents at runtime, the same layer as Okta's Agent SSO. The move that ages well is to keep the commerce protocol, the payment rail, and the identity layer swappable rather than betting the build on a single winner. It is the same decoupling that routing through a harness bought us on the model side: keep the agent logic independent of which standard a given merchant speaks.

In summary, A person issues a credential that constrains what, where, and how much an agent can spend, the agent presents it, the endpoint verifies the constraints and the agent's identity, and the rail settles. Build that boundary in from the start and you are aligned with where the platforms and card networks are already heading: Agentic eCommerce.

Frontier Pricing Dropped Again

Two months ago I wrote about a pricing war compressing frontier capability toward commodity. September kept the trend going. Anthropic released Claude Fable 5.1 on September 1 as its top model for long-running agents, coding, and research. Claude Opus 5.5 arrived September 22 at $4 per million input tokens and $20 per million output tokens, roughly 20% below Opus 5, and Anthropic says it runs about 40% cheaper than Opus 5 on typical workloads while matching Fable 5.1 on most work. It tops the Artificial Analysis Intelligence Index at around 57 to 58. Claude Sonnet 5.5 followed on September 28, holding at $2 and $10.

OpenAI positioned GPT-6 Astra as its frontier model for coding, computer use, and science, available through its API and on AWS, with the lower-cost GPT-6.1 Sol introduced around DevDay on September 29 at roughly one-fifth of Astra's token price.

AWS Rebuilt the Agent Runtime

AWS has shipped a new AgentCore Runtime that changes the economics of running agents, not only their feature set. Two things stand out together. Elastic memory reclaims what a session is not using instead of holding allocation at the session's peak, so the infrastructure follows the workload. Snapshot-based starts bring the P75 cold start to roughly two seconds for container images between 200 MB and 2 GB, down from about 5 to 30 seconds on the previous runtime. Scale-to-zero and hardware-enforced session isolation stay in place.

AWS added more in the same stretch: interactive shell sessions that live inside the isolated session, with up to ten shells per harness; AgentCore Evaluations extended to TypeScript frameworks including Strands, LangGraph, and the Vercel AI SDK; and a managed consent portal for AgentCore Identity that lets an agent get OAuth access to a third-party service without a team building its own callback plumbing.

(Source: https://aws.amazon.com/blogs/machine-learning/the-new-agentcore-runtime-elastic-optimized-and-consistently-fast-starts/)

A Great Focus on Agent Identity

At Oktane, Okta launched Agent SSO, giving agents managed sign-on the way employees already have it. The theme carried into October as vendors shipped per-agent credentials an agent must present at runtime. Much needed, and we look forward to incorporating this into our security options for agentic harness security.

The adoption data explains the rush. G2's 2026 AI Agents Report found that 66% of organizations now run agents in limited or scaled production, with adoption up 25% year over year. In mature, high-volume workflows, four of five teams reported cycle-time improvements around 11%, and half reported speed gains of up to 25%.

(Source: G2 2026 AI Agents Report, via Diginomica)

Once two-thirds of enterprises run agents in production, the hard question shifts to identity: who the agent is and what it is allowed to touch. The Muse block was the same question asked from the outside, by companies deciding which agents to let near their systems. Inside your own walls it is an access and auditability problem, and the market is now funding and shipping it as its own category rather than leaving it for each team to solve by hand.

Bottom Line

Now that we have seen agents locked out of the web's biggest storefront, cheaper frontier models, a runtime that makes always-on agents affordable, and an identity layer arriving as a product. The through-line is control: who the agent is, what it is allowed to reach, and whether you can see and revoke that access. The agent you cannot account for is how a program stalls between pilot and production, or how it gets turned away at someone else's door. If your team is running agents on AWS, the rebuilt AgentCore Runtime and AgentCore Identity belong on this quarter's evaluation list, paired with the governance and cost controls that keep an agent inside the limits you set. That is the work UTurn's Realization Factory and AI Production Build Lab are built to do. The alternative is finding out during an audit who authorized the agent, which is a worse time to learn it.

Alec MacEachern is Vice President of AI at UTurn Data Solutions, an AWS Premier Tier Services Consulting Partner based in Chicago. Over the past decade, he has held roles at NVIDIA, AWS, and Microsoft, helping organizations design, build, and scale AI solutions across a wide range of industries and platforms. Today, Alec brings that cross-platform experience to helping enterprises navigate cloud migration, modernize data foundations, and adopt production-ready generative and agentic AI solutions.