- Copilot Studio agents (15 Oct) get their own Entra ID and mailbox and can start on an email or a schedule. Preview only, with real limits on memory privacy.
- ServiceNow EmployeeWorks (15 Oct) lets staff build background "Productivity Routines" without code. ToolSmith builds automations from plain language.
- NVIDIA Nemotron 3.5 (16 Oct) is open. NVIDIA says its Bolt model fits in 24GB, so an agent can run on one local graphics card.
- VideoGen Model 1 (20 Oct) makes multi-shot videos up to 10 minutes with consistent characters. Rollout starts with private partners.
Copilot Studio agents get their own ID, mailbox and memory
Microsoft announced that account-based agents built in Copilot Studio get their own Entra ID, a mailbox and a place in Microsoft 365, like a colleague. They can start work when an email arrives or on a fixed schedule, and they remember things between conversations.
The preview has clear limits. The agents cannot respond outside the company's tenant, cannot read documents labelled Confidential or Internal Only, and cannot make targeted edits to existing documents. Email is limited to conversations inside the tenant with up to 250 people. Shared memory in the preview has "no user privacy boundary" and no opt-out, and per-user memory disappears after 28 days without interaction. Access runs through the Frontier programme and needs Microsoft 365 E5 with the Agent 365 add-on, or E7, plus a Frontier for AI Teammates licence per agent. Copilot Credits are billed on usage.
Your angleAn agent with its own mailbox needs onboarding like a new hire: which folders it sees, which emails start it, who reviews its work. Because shared memory has no privacy boundary in the preview, part of the job is telling the client what not to put in front of it yet.
Source: NeoTeo
ServiceNow EmployeeWorks: staff build their own agents
ServiceNow introduced the third generation of its employee-centric AI harness behind EmployeeWorks. Employees without coding skills can build Productivity Routines: they describe a repeatable task, pick the tools and deploy it. The routine then runs in the background and follows the same steps every time, with governance built in.
For developers there is ToolSmith, an environment where you describe the automation in plain language and it does the build. ServiceNow says it was trained on more than 75,000 APIs and can read logs to suggest fixes and new workflows. All of it is in controlled availability for now.
Your angle"Anyone can build a routine" still means someone has to decide which routines are worth building. A short workshop that maps a team's weekly repeat tasks and turns the top three into routines is a concrete, sellable package, and the same method works on any agent platform.
Source: ServiceNow press release
NVIDIA Nemotron 3.5: open agent models that run locally
NVIDIA launched three open Nemotron 3.5 models for enterprise agents. Super VL is the largest; NVIDIA says it delivers up to 5x the throughput with about a fifth of the parameters and was trained on 26 trillion tokens. Bolt 74B-A9B has 9 billion active parameters and, according to NVIDIA, fits in 24GB with 2-bit quantisation, so it runs on an RTX card or DGX Spark. NVIDIA reports it ranks first on PinchBench and second on BrowseComp. A smaller Bolt E4B (4GB) is due on 23 October. CrowdStrike, Atlassian and Salesforce are named as users.
Mitch Ashley of Futurum summed up the trade-off: "The inference bill shrinks, and the governance surface grows."
Your angleClients who will not send data to a cloud model (law firms, clinics, accountants) now have a realistic local option. Setting up an agent on their own hardware, testing it on their real tasks and writing down who may change it is a service with a clear scope. Benchmarks are NVIDIA's own, so test before you promise quality.
Source: Techstrong.ai
VideoGen Model 1: multi-shot videos up to 10 minutes
VideoGen unveiled Model 1, its first video generation model. It makes multi-shot videos of up to 10 minutes, takes up to 16 reference images and keeps characters and locations consistent. A "thinking harness" writes the script first. VideoGen says it is more than 50% cheaper, based on its own figures. Videos carry invisible watermarks. The rollout starts with a limited group of private partners, and the company opened sign-ups for $1 million in free credits on 20 October.
CEO Anton Koenig called it "a major step toward AI becoming part of everyday production," from film pre-visualisation to full-length YouTube videos and ads.
Your angleConsistent characters mean a client can have one recurring presenter across explainers, onboarding and ads. Building that character, a script template and a monthly batch of videos is a retainer, not a one-off. Get the client's sign-off on how AI video is labelled before anything is published.
Every launch makes agents easier to build. Deciding what they should do is still the job.
Overview: four launches, four services
Based on the announcements linked above. Most of this is preview or limited access, so check the vendor page before you advise a client.
| Copilot Studio | EmployeeWorks | Nemotron 3.5 | VideoGen Model 1 | |
|---|---|---|---|---|
| Announced | 15 Oct 2026 | 15 Oct 2026 | 16 Oct 2026 | 20 Oct 2026 |
| What it is | Agents with own ID and mailbox | No-code background routines | Open models for agents | Long-form video model |
| Access | Frontier preview, E5 + add-on or E7 | Controlled availability | Download now (Bolt E4B 23 Oct) | Private partners first, waitlist |
| Main limit | Tenant only, shared memory not private | ServiceNow customers only | Benchmarks are vendor claims | Cost claim from own figures |
| Service to sell | Agent onboarding and permissions | Routine design workshop | Private, local agent setup | Recurring video with one character |
Before you sell any of these: 4 checks
- Can the client actually get it? Three of the four are preview or limited access. Do not sell a tool the client cannot buy yet.
- What data does it see? Mailboxes and shared memory are powerful and risky. Decide what stays out.
- Who approves the risky steps? Sending, paying and publishing need a named human. Our 7-check security checklist covers the rest.
- How do you measure it? Agree on one number before you start, such as hours saved per week, and check it after a month.
An ID and a mailbox mean agents get onboarded like staff. That onboarding is work.
Open models on one graphics card open the door to clients who said no to the cloud.
Consistent characters turn AI video from a one-off trick into a monthly deliverable.
AI Money Class #2 with Jarvy: your first AI client in 7 days (swipe →)
Get the next briefing in your inbox
One short email when there is AI agent news that matters for people who sell to businesses. Sources included, no hype.
FAQ
What are the new Copilot Studio agents with their own mailbox?
Microsoft announced on 15 October 2026 that account-based agents built in Copilot Studio get their own Entra ID, a mailbox and a place in Microsoft 365. They can be triggered by an email or a schedule. In the preview they cannot respond outside the tenant, cannot read documents labelled Confidential or Internal Only, and shared memory has no user privacy boundary. Access runs through Microsoft's Frontier programme with E5 plus the Agent 365 add-on or E7, plus a per-agent licence.
Can I run an AI agent model locally for a client with sensitive data?
NVIDIA released Nemotron 3.5 on 16 October 2026 with open weights. According to NVIDIA, the Bolt 74B-A9B model fits in 24GB with 2-bit quantisation, so it can run on an RTX card or DGX Spark. Local models keep data on your own hardware, but you still need to test quality on the client's real tasks.
Are these tools available to everyone?
Not yet. The Copilot Studio agents are in preview through the Frontier programme, ServiceNow's new features are in controlled availability, and VideoGen Model 1 rolls out to a limited group of private partners first with a waitlist. NVIDIA's Nemotron 3.5 Super VL and Bolt models are downloadable now; Bolt E4B is due on 23 October.
Recorded training · real agent screens
Watch a team of AI agents do real client work
In the free training you will:
- See agents handle leads, follow-up, content research and spreadsheets, step by step.
- Learn how to use it for your own business or whitelabel it for clients.
- Leave with a next-step plan you can start the same week. No coding required.
Free training. No income guarantee: results depend on your own work and clients.
What you learn to deliverReady to start this week?
Skip the wait: get the full course
€497 one-time · instant login after paying
- 6 modules: build agents for leads, follow-up, research and reporting
- Price calculator + proposal and outreach templates
- Progress tracking, lifetime access, no coding needed
Not sure yet? Watch the short video first · compare all 3 options
Secure checkout via Stripe (+2% service fee). No income guarantee: results depend on your own work and clients.
All sources
- NeoTeo: Microsoft announces persistent agents for Copilot Studio (16 Oct 2026)
- ServiceNow: Third generation of its employee-centric AI harness (15 Oct 2026)
- Techstrong.ai: NVIDIA launches three open AI models to power enterprise agents (16 Oct 2026)
- GlobeNewswire via The Manila Times: VideoGen unveils Model 1 (21 Oct 2026)
- AI Agents Directory: AI agents news brief (17 Oct 2026)