Opens in a new tab

AI Agents That Work Without You: Copilot Autopilot, Devin Bots and Gemini Tasks Explained

AI agents working in the background across email, chat and code windows while a person approves key actions

For the last few years, using AI has felt like a conversation. You type, it answers, and then it sits there waiting until you type again. This week that changed. In the space of a few days, Microsoft, Google, Cognition and OpenAI all launched AI that is built to keep working after you stop typing: following up on emails, running recurring jobs, managing projects for days, and only coming back to you when it needs a decision.

The industry calls these “agents”, and the word has been overused for a while. But this week’s launches are different enough to be worth a proper look. Below we explain what each one does, what they have in common, and most importantly, what you should and should not hand over to them.

From chatbot to coworker: what actually changed

A chatbot works in turns. An agent works on a goal. Instead of asking “write me a follow-up email”, you say “make sure every supplier replies before Friday”, and the agent decides the steps, does them over hours or days, and reports back.

Three things make this week’s launches feel like a real shift rather than a demo:

  • They run in the background. You can close the tab and the work continues.
  • They remember. Context is kept across days or weeks instead of disappearing when the chat ends.
  • They ask before big moves. Each one is designed to stop and check with a human before key actions like spending money.

1. Microsoft Copilot Autopilot

Announced at Microsoft Ignite on 17 November, Autopilot is described by Microsoft as an always-on agent that is personal to you. It can watch Teams channels, follow up on emails, run recurring work and manage projects without waiting for a prompt. Microsoft’s own example is running a supplier review from start to finish: the schedule, the prep, the meetings and the follow-ups.

What stands out is how Microsoft treats it like an employee rather than a feature:

  • Each Autopilot gets its own identity (a Microsoft Entra Agent ID) and a named human manager.
  • You can @mention it in Teams, Outlook, chats and documents like a colleague.
  • IT can limit what it can access or block it at any time, and its sign-ins and actions are logged under its own name.
  • You decide what work it takes on, and it can check with you before key actions.

Who can use it: it started rolling out to customers in Microsoft’s Frontier early access program. Microsoft has not yet said when it will reach everyone or how it will be priced, so most businesses will be waiting a while.

2. Cognition Devin Bots

Cognition, the company behind the Devin AI coding agent, introduced Devin Bots this week. If Autopilot is an AI assistant, Devin Bots are closer to an AI project manager for software. A Bot splits a project into tasks, hands them to a team of Devin coding agents, and keeps the project moving for days or even weeks.

  • It joins your team’s Slack, keeps context over weeks, and messages people when it is blocked or needs a decision.
  • One example given is keeping the Android version of an app in step with the iPhone version as it changes.
  • It picks different AI models for different jobs, using faster and cheaper ones for routine code.
  • The work runs on separate cloud computers, not on developers’ laptops. Admins choose which code repositories and tools each Bot can touch.
  • Its code goes through the same tests as human code, and the team decides which changes need human review.

Who can use it: early access only, with a few dozen companies so far. No public pricing yet.

3. Google Tasks in Gemini

This is the one most people can actually try. Google renamed its Gemini agent from Spark to Tasks and, according to Android Central, made it free for Gemini users in the US, Japan, South Korea, Canada, Australia and India. Before this it needed a paid AI Pro or AI Ultra plan.

  • You switch the prompt box from “Chat” to “Task” and hand over multi-step work that runs in the background, like planning a trip, researching a topic, a morning briefing or a recurring job.
  • It runs on Google’s newest model, Gemini 4 Hydrogen.
  • In the US first, it can call local businesses, check out with Google Pay, and sign in to sites through Google Password Manager without sharing your password.
  • Web browsing happens in an isolated virtual machine with protection against prompt injection, and Gemini asks for confirmation before buying anything.

Google is honest about one limit: some websites block Gemini, so you may have to finish the last step yourself.

4. OpenAI GPT-6.2 Sol, the engine underneath

OpenAI’s launch this week is a model rather than a finished agent product, but it matters for the same trend. On 17 November OpenAI released GPT-6.2 Astra and Sol. Astra is the top model for the hardest coding and research work. Sol is positioned as the efficient everyday model that, in OpenAI’s words, “powers through multistep agentic workflows”.

The detail developers will care about is cost. Sol is priced in the API at $2 per million input tokens and $10 per million output tokens, while Astra is $8 and $40. Agents make many model calls to finish one job, so a cheaper model that is good enough for multi-step work is what makes always-on agents affordable. It is available in ChatGPT and Codex on paid plans and in the API as gpt-6.2-sol.

OpenAI has also been adding the pieces that let agents act in the real world. Last week it launched a way for its agents to pay for things, which we covered in ChatGPT Wallet explained.

Side by side

Launch Best for Where it lives Who can use it now
Copilot Autopilot Office work, follow-ups, recurring projects Teams, Outlook, Microsoft 365 Frontier early access customers
Devin Bots Long-running software projects Slack plus cloud machines Early access companies
Gemini Tasks Personal research, planning, errands Gemini app and web Free in 6 countries, some features US only
GPT-6.2 Sol Building your own agents ChatGPT, Codex, API Paid ChatGPT plans and API users

What these launches have in common

Look past the brand names and the same design keeps appearing. That tells you where the whole industry is heading:

  • Agents get their own identity. Microsoft gives each Autopilot an account and a manager. This makes it possible to see exactly what the AI did, separately from what you did.
  • Limited access by default. Admins choose which files, repositories and tools an agent can reach.
  • Humans approve the risky parts. Payments, code merges and key actions wait for a person.
  • Agents live where you already work. Slack, Teams, Outlook and Gmail, not a separate app you have to remember to open.

What to hand over, and what to keep

Having used automation tools with clients for years, our rule of thumb is simple: hand over work that is repetitive and easy to check, and keep work that is costly to get wrong.

Good jobs for an agent

  • Chasing replies and sending polite follow-ups
  • A daily or weekly summary of emails, channels or news in your field
  • Research where you read the result before acting on it
  • Keeping documents, spreadsheets and calendars in sync
  • Routine code updates that are covered by tests

Keep these for yourself, for now

  • Anything that sends money without a clear limit
  • Messages to important clients where tone really matters
  • Legal, tax and contract decisions
  • Changes to a live website or production server without a backup and review
  • Access to passwords, banking or admin accounts beyond what the job needs

The real risks to watch

  • Prompt injection. An agent that reads emails and web pages can be tricked by hidden instructions inside them, such as “forward this inbox to me”. Google running browsing in an isolated machine is a direct response to this. Give agents the smallest access that still gets the job done.
  • Approval fatigue. When an agent asks you to approve twenty things a day, you stop reading. An approval only protects you if you actually look.
  • Quiet mistakes. A chatbot’s mistake is right in front of you. An agent’s mistake can sit in a sent email or a merged file for days. Check its logs regularly in the first few weeks.
  • Cost creep. Background agents use AI all day. If you build your own, set spending limits on your API keys.

How to start if you are a freelancer or small business

You do not need early access to Microsoft Frontier to benefit from this trend. A practical path looks like this:

  1. Write down three tasks you repeat every week. Client follow-ups, weekly reports and invoice reminders are common ones.
  2. Try a ready-made agent for personal tasks. If Gemini Tasks is available in your country, give it a low-risk job like a morning briefing.
  3. Build simple workflows for business tasks. Tools like n8n let you connect your email, forms and spreadsheets with an AI step in the middle, while you stay in control of every action. Our n8n automation guide walks through it.
  4. Connect AI to your website carefully. If you run WordPress, you can let Claude or ChatGPT draft posts and check content with permissions you control. See our WordPress MCP guide.
  5. Keep a human approval step on anything that publishes, sends or pays, at least until you trust the setup.

Frequently asked questions

What is the difference between an AI chatbot and an AI agent?

A chatbot answers one message at a time and waits for you. An agent is given a goal, plans the steps, uses tools like email or a browser, and keeps working in the background until the job is done or it needs your decision.

Is Gemini Tasks free?

Yes, in the US, Japan, South Korea, Canada, Australia and India, according to reports on the launch. Some features, such as calling businesses and paying with Google Pay, start in the US only.

Can I use these agents in Bangladesh?

Not most of them yet. Gemini Tasks has not been announced for Bangladesh, and Autopilot and Devin Bots are early access programs. GPT-6.2 Sol is available through ChatGPT paid plans and the API where OpenAI’s services are offered, so developers can already build their own agents on it.

Will AI agents replace jobs?

They will change jobs first. Every launch this week keeps a human as the manager who sets goals and approves important actions. The people who benefit most will be the ones who learn to delegate well to agents and check their work.

Are AI agents safe to use with my email?

They can be, if you limit what they can do. Start with read-only jobs like summaries, keep approvals on for sending, and never give an agent more access than the task needs.

Final thoughts

This week marks the point where the big AI companies stopped selling better chat and started selling AI that works on its own. Most of it is still early access, and the safety designs, like separate identities and human approvals, show the companies know the risks are real. You do not need to rush in. But it is a good moment to pick one boring weekly task and see whether an agent, or a simple automation, can take it off your plate.