This week the agent ecosystem stopped looking like a collection of demos and started looking like an infrastructure stack. OpenAI is training frontier models offline. Anthropic has locked in up to one million Google TPUs. Google Cloud launched a dedicated compute engine for agents. AWS open-sourced a runtime that can park 75,000 idle agents in 8 GB. Anthropic donated Agent Skills to the Linux Foundation, NVIDIA expanded its world-model tools, and Cursor handed agents the mouse and keyboard. For SMEs, the signal is clear: the runtime layer underneath AI agents is now real, and that changes what a proof of concept can prove.
What happened this week
OpenAI published how it secures frontier training: the most capable reinforcement learning runs now live in an air-gapped offline world. Models still browse and act on content, but that content is a cached copy, not the live internet. OpenAI red-teamed the setup with its own frontier agents, including the unreleased GPT-6 Astra Cyber, to make sure the walls hold.
Anthropic announced a Google Cloud expansion worth tens of billions of dollars, covering up to one million TPUs in 2026 and bringing more than a gigawatt of capacity online. It also reaffirmed Amazon as its primary training partner through Project Rainier, so the company is spreading compute across Google, Amazon and NVIDIA rather than betting on one provider.
Google Cloud introduced Agent Compute Engine at AGNTCon + MCPCon on October 22. Each agent gets its own hardware-isolated microVM with a dedicated guest kernel. Idle agents are checkpointed and restored, and copy-on-write memory sharing lets many agents run from one OS template. The goal is security and cost at scale.
AWS released Infinity, an Apache 2.0 research runtime from the Hydro project with UC Berkeley. The insight is simple: agents spend most of their time waiting. A suspended agent costs about 100 KB of bookkeeping. AWS benchmarked 80,000 agents across 3.2 million completions inside one large instance.
Anthropic donated Agent Skills to the Agentic AI Foundation under the Linux Foundation. The standard is already integrated by more than 70 platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code, with roughly one million skills published and nearly 280 million hours of agent compute.
NVIDIA expanded Cosmos with Data Factory, framework recipes for fine-tuning, RL, distillation and quantization, all packaged as agent skills. Cursor closed the loop by letting its agents control desktop applications, including the iOS Simulator, to test whatever they just built.
Why this matters for SMEs
Small businesses do not need a million TPUs or a frontier model lab. They need three things: a way to run an agent safely, a way to connect it to the tools they already use, and a way to measure whether it actually saves time or money. The week’s announcements hit all three.
Isolated microVMs and offline training environments mean enterprise-grade security is becoming table stakes. Cheap suspension and high concurrency mean an agent that sits idle 95% of the time no longer needs its own dedicated server. Open standards like Agent Skills mean an agent built for one tool can talk to others without a custom integration every time.
The result is that a focused POC no longer has to be a toy. It can read a real inbox, update a real spreadsheet, book a real calendar slot, or call a real API, and it can do so inside a sandbox that is cheap enough to throw away if it fails.
What we recommend testing first
Instead of asking which agent platform to buy, pick a single workflow that is repetitive, measurable, and safe to get wrong a few times. Common starting points we see work:
- Scheduling and reminders — an agent reads incoming requests, checks availability, and proposes slots.
- Draft generation — an agent prepares first-pass quotes, invoices, or customer replies from structured data.
- Status updates — an agent pulls information from one system and writes a concise summary into another.
Keep the scope narrow enough to measure in a week. Set one success metric: time saved, errors avoided, or customer response time. Run it against real data, but keep a human in the loop. That is how you find out whether the runtime layer actually improves your business before you scale it.
How Team19 thinks about it
At Team19 we build POC projects for SMEs that want to see what agents can do on their own data and processes. We do not start with a platform pitch. We start by asking what takes up too much time, where handoffs break, and what a small team could stop doing if an agent handled it.
The October infrastructure news makes those POCs faster, safer, and cheaper to run. If you want to test one workflow with a real agent in your real stack, [email protected] is the fastest way to talk to a human on our team.