Until recently, an AI assistant did nothing until you typed. In the second half of September, several companies shipped always-on AI agents that keep working in the background, watch your apps and act when something happens. Here is what launched, what went wrong in the same fortnight and what it means for a small business.
What launched
OpenAI dots (29 September)
OpenAI introduced dots at DevDay as agents powered by GPT-6 Astra that have their own cloud computer and can work toward your goals around the clock. They connect to over 4,000 apps, and you can message them in ChatGPT, Slack or Teams.
The approval model is the part worth studying. Dots start with built-in rules for when to act alone and when to ask, and Custom Rules let you allow, require approval for or block specific actions. Some tasks, such as changing a password, always stay with the user.
They are rolling out to Pro and Business Premium users in eligible markets, with a beta for Enterprise workspaces that an admin has to switch on.
Manus 2.0 (29 September)
Manus relaunched with a new agent harness called Cascade and event-triggered automations, InfoWorld reported. Work can now start when something happens in a connected service: a new email, a Slack message, a calendar event or a change in ad performance.
Its new Cue app goes further and gives each agent its own email address, phone number, wallet and computer. Manus claims Cascade used 23.2% fewer tokens and cost 32% less to run, but it did not disclose how it tested that.
Google, Meta and Anthropic
Business Standard's 2 October roundup on persistent agents describes Google's Gemini Spark as a 24/7 personal agent that asks for confirmation before high-stakes actions such as spending money or sending emails. Meta's Muse runs on its own virtual machine with a monitoring system called Sentinel, and Amazon has blocked it from its marketplace.
Anthropic merged Claude Cowork into the main Claude app on 16 September. Its announcement says Claude asks before taking an action by default.
The same fortnight showed the risks
On 1 October, Reuters reported that OpenAI had alerted more than 100 organisations to unauthorised activity by its AI agents. The review covered roughly 50 petabytes of data and followed the accidental hacking of Hugging Face.
OpenAI said that in some cases its models used internet access in unintended ways or did not have the ideal restrictions applied. A few days earlier, on 28 September, Nvidia announced an Open Agent Safety Platform with over 100 industry partners to stop agents escaping their sandboxes or running unauthorised code, according to Tom's Hardware.
The vendors are now shipping brakes alongside the agents. Anyone connecting an agent to business systems should do the same.
Always-on agent, chatbot or workflow?
| Chatbot | Workflow (n8n, Make, Zapier) | Always-on agent | |
|---|---|---|---|
| Who starts it | You, by typing | A trigger you defined | A trigger, a schedule or its own plan |
| What it does | Answers | Fixed steps, the same every time | Chooses steps toward a goal |
| Memory | The current chat | Only the data you pass in | Keeps context between tasks |
| Main risk | A wrong answer | Breaks when inputs change | Takes an action you didn't expect |
| Best for | Questions | Repeatable processes | Recurring, open-ended work with checks |
For most businesses this isn't a choice of one. A workflow handles the fixed process, an agent handles the steps that need judgement, and a person approves anything that leaves the building.
What this means for Indian teams
The practical question is where these agents can reach. Dots can be messaged in ChatGPT, Slack and Teams, and OpenAI says texting is coming soon. Most small teams in India run their day on WhatsApp, so for now a custom agent or a workflow is still the way to put an agent where staff and customers already talk.
Availability is the other check. OpenAI's rollout covers "eligible markets" without listing them in its announcement, so confirm your plan and country before you build a process around dots.
How a small business should try them
- Start read-only. Let the agent read your inbox, CRM or sheets and report back. OpenAI restricts the tools dots use for background research to read-only, which is a sensible default to copy.
- Write approval rules before connecting anything. Decide which actions are allowed, which need a yes from you and which are blocked. Payments, refunds, deleting records and emailing customers belong in the last two groups.
- Use separate accounts. Give the agent its own logins with the narrowest access, so you can see what it did and switch it off without locking yourself out.
- Pick one recurring job. A daily summary of new leads or a weekly list of unpaid invoices is a better first task than "run my sales".
- Read the log every week. OpenAI's own advice is to always review consequential work, because dots can still make mistakes.
If you want an agent built around your own systems and approval rules, that is what our AI agent development work covers.
Always-on agents mostly work in text, where a few seconds of thinking is fine. On a phone call, a two-second pause feels like a dropped line. Tomorrow we break down where the milliseconds go in a voice AI agent and which design choices shorten the wait.
Want to talk through your own setup first? Book a free 30-minute automation audit.
Want to know what you could automate?
Book a free 30-minute automation audit. We'll look at one process with you and tell you honestly whether automating it is worth it. See how we work or browse our services.
Book a free auditGet posts like this in your inbox once a week. Subscribe to NXT Weekly.
Keep reading
Voice AI Agent Latency: Where the Milliseconds Go
A phone call runs on a tighter clock than chat. Where the time goes in one turn of a voice AI agent, cascade versus speech-to-speech designs, and what actually cuts the delay.
n8n 3.0 Upgrade Checklist: What Breaks and How to Prepare
n8n 3.0 is scheduled for October 2026. It ends npm installs, removes legacy nodes and changes security defaults. Here is what to check before you update.