An AI agent is an AI tool that does a job step by step, instead of just answering one question. I opened every source below myself and read it before writing. When a number comes from the company selling the product, I say so.
Microsoft gives Copilot Studio agents their own accounts
Microsoft announced on 15 October 2026 a preview of agent accounts, shared memory and triggers for Copilot Studio. I read NeoTeo's report of it, published the next day. This is one source, so these are Microsoft's claims as NeoTeo reports them.
An account-based agent gets its own Microsoft Entra ID identity, a mailbox and a presence in Microsoft 365. A trigger can be an email or a set schedule. It starts the agent without anyone typing a new prompt.
There are limits. These agents cannot answer people outside the company's tenant. They cannot read documents labeled "Confidential" or "Internal Only". Email replies stay inside conversations of up to 250 people.
The memory part is worth a careful read. In the preview, an account agent's memory is shared among the people who work with it. There is no private boundary per user and no opt-out. NeoTeo also says it needs Microsoft 365 E5 with the Agent 365 add-on, or E7, plus a separate licence per published agent.
My take: Microsoft makes and sells this, so the benefits are Microsoft's description. NeoTeo calls the account "an identity, not a blank check". I like that, but shared memory with no opt-out is something I would test before putting private work in it.
NVIDIA releases open models built for running agents locally
NVIDIA launched three open models in its Nemotron 3.5 family, according to Techstrong.ai on 16 October 2026. Again this is one source, and most numbers are NVIDIA's own.
The biggest, Nemotron 3.5 Super VL, has 120 billion parameters and is made for jobs like document processing and working with business software. Its weights are openly available, so a company can train it further on its own data and run it in its own systems. It is on Hugging Face, OpenRouter, ModelScope and NVIDIA NIM. The article does not give the licence terms.
The middle model, Bolt 74B-A9B, uses 9 billion of its 74 billion parameters at a time and, with two-bit quantization, fits in 24GB of memory on NVIDIA RTX and DGX Spark machines. The smallest, Bolt E4B 4GB, is due on 23 October. NVIDIA claims Super VL has up to five times the throughput of competing models in its class, and that Bolt 74B ranks first on PinchBench among comparable open models.
Techstrong says CrowdStrike is customizing Super VL, Atlassian is exploring it for search across Jira and Confluence, and Salesforce is adapting it for government workflows.
My take: NVIDIA sells the hardware these models run best on, so read the speed and ranking claims with that in mind. The analyst Mitch Ashley of The Futurum Group put the trade-off well: "The inference bill shrinks, and the governance surface grows." Local agents can slip past the cloud gateways IT uses to watch AI use.
American Express shares a way to prove what you approved
American Express published its Intent Intelligence framework for agentic commerce, which means AI agents buying things for you. I read two reports: The Paypers (14 October 2026) and Trew Knowledge's weekly recap (16 October 2026). Amex's own page would not load for me, so everything here is how those two describe Amex's claims.
The idea: the agent writes down what you asked for, such as the merchant and your budget. You verify who you are and approve it. That creates an Intent ID and a Proof of Intent Token. The token is then used to get a payment credential limited to what you approved.
Trew Knowledge gives Amex's example: if an agent tries to buy Adidas shoes when you asked for white Nikes in size 9, the purchase can be blocked. Amex says the card issuer sees the approved intent before the purchase, not after. The framework is reachable through Amex's ACE Developer Kit, and Amex says it is working with EMVCo and supports the x402 Foundation.
My take: Amex runs a payment network, so it benefits if people trust agent payments. Still, a spending limit that the agent cannot go beyond is the right kind of idea. I have not seen independent tests of how well it works.
Vertafore launches two agents for insurance paperwork
Vertafore, an insurance software company, made two AI agents generally available, according to FinTech Global on 16 October 2026.
The Policy Intake Agent reads documents like declarations pages, binders and PDFs, matches the details to fields in the agency system, and flags anything new, changed, missing or contradictory. The Policy Check Agent compares carrier policies with the agency records and suggests corrections. In both cases staff approve the changes before the agent updates anything. The check agent also logs each review so there is a record.
Vertafore says its early trials showed up to 87% less time on policy intake and up to 90% less on policy checking. FinTech Global says these are from Vertafore's early trials and customer reports, and it did not verify them.
My take: The numbers come from the company selling the agents, and "up to" means the best case. What I do like is the design: the agent suggests, a person approves. That pattern showed up in several stories this week.
What I take from this week
- Agents are getting identities. Microsoft's agent accounts and Amex's Proof of Intent Token are both ways to say who or what is allowed to act.
- Human approval is still the default in the business tools. Vertafore's agents wait for staff to approve, and Amex's framework starts with the customer approving.
- Running agents locally gives more control over data, but the NVIDIA story is a reminder that it also makes them harder for IT to watch.
Most of the numbers this week come from the companies themselves, and I found no independent tests. Treat them as claims until someone outside checks them. I will keep reading the sources myself and tell you which ones I could not open.
Sources
- NeoTeo: Microsoft announces persistent agents for Copilot Studio
- Techstrong.ai: NVIDIA Launches Three Open AI Models to Power Enterprise Agents
- The Paypers: American Express publishes framework for agentic commerce intent
- Trew Knowledge: AI This Week: Agents Get an Identity
- FinTech Global: Vertafore launches AI agents to automate insurance policy checks