The weekly note

The week in agentic AI: November 9-14, 2026

Arkatai

This week brought three moves that point at the same spot: who controls the channel an agent lives in, and who answers for what it does once it runs longer without direct supervision. One survey put numbers on how many companies have moved from testing to running agents. Two labs, on the same day, expanded what an agent can do in front of a customer and how long it can work unsupervised. Here is what each piece changes for a company that already operates agents.

What happened

Zapier released the 2026 update of its annual survey of business leaders, and the picture has genuinely shifted. 88% of the leaders surveyed say their organization is using or testing AI agents, up from 72% in 2025. The real change is in the mix: production use climbed from 40% to 71%, while testing dropped from 32% to 17%. 90% say they are certain or likely to increase investment next year, up from 84% last year. For a company still running its agent “as a pilot,” the number that matters isn’t the 88% adoption figure, it’s the 17% still testing: the middle stage is emptying out, and most companies land in production, not in a drawer.

OpenAI launched “Businesses in ChatGPT” in beta, with two pieces: Business Hub and Business Agents. Companies can claim and verify their profile inside ChatGPT and see how many mentions and clicks they get. On top of that profile, eligible businesses can publish their own agent that answers questions about their products in a conversation kept separate from the main chat. For now it’s limited to US ecommerce and online-service businesses, free at launch, and OpenAI says it won’t share users’ conversations with the business. For a company running customer service through agents, the interesting part isn’t the feature, it’s that the customer-facing agent now lives inside a third party’s platform, under whatever visibility, moderation, and future advertising rules that third party sets. That’s exactly the question I cover in provider or deployer: who answers when the agent acts isn’t automatic, it depends on where the agent lives.

Google announced Gemini 4 Carbon, its most capable model for cybersecurity and for the agentic work that pushes a model hardest. Google itself is clear about the split: Argon remains the choice for most software engineering and knowledge work, and Carbon is reserved for tasks that test a model’s limits, with long-horizon agentic workflows and extended verification loops. It’s available through the API in Google AI Studio and Gemini for business at an introductory price of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 on November 30. Google AI Ultra subscribers also get it in Antigravity, where Google previewed a feature called “/swarm” that splits a task across several coordinated agents working in isolated workspaces. The longer a leash an agent gets without anyone watching, the more weight falls on the question of what it can do on its own and what it has to escalate, which is the territory of AI agent permissions and controls and autonomous AI agents: autonomy isn’t a fixed property of the model, it’s a business decision made per type of action.

How to read it from the operation

All three pieces point at the same transition seen from different angles. Zapier’s survey confirms with numbers what was already visible: the question that actually matters to a board is no longer whether piloting an agent is worth it, it’s whether the process in front of them is ready to move to production with an owner who answers for its result. OpenAI’s and Google’s launches show where the lab competition is actually being fought: who controls the customer-facing channel, and who can run longer without a human in the loop. Neither changes the underlying question for a company that operates agents: a more capable model or a more convenient channel doesn’t replace the work of defining each agent’s mandate, where it stops and escalates to a person, and who reviews its evaluations when the model changes. I cover that in the essay agent governance. The difference between buying a platform that answers for its uptime and running a process whose result someone guarantees is still the same difference as always, only now it gets decided over models given an ever-longer leash without anyone watching.

What to watch

On November 30, 2026, Gemini 4 Carbon’s introductory API pricing rises from $2/$10 to $4/$20 per million input/output tokens, according to Google’s own announcement. Anyone evaluating the model for a high-volume use case has less than three weeks to measure the real cost before that number changes.