Browsing a website, running a scheduled job or sending a message is no longer the difficult part of building an AI agent. Hermes Agent can do it, OpenClaw can do it, and Codex can increasingly do it too.

What still takes work is turning all that capability into something you can hand a real job to without spending half your time configuring and managing the agent.

SpaceXAI and Cursor put Grok Bot into early beta on August 11. You create named teammates, hand them work, and they finish it inside your real tools while your laptop is shut.

While everyone was sharing clips of Grok Bot clearing inboxes this week, I sat down, read the documentation and tried it briefly instead. The demos are impressive, but the docs are more useful, because they tell you what you are agreeing to.

One Topic: New Personal AI Agent – Grok Bot

What exactly is Grok Bot, in simple terms

Think of Grok Bot as a persistent AI worker with access to a cloud computer.

You create a Bot, give it a name and a short job description, then message it the way you would text a colleague. Behind that chat window sits a cloud Linux machine with a browser, a terminal and a file system, running whether or not your laptop is on. The Bot signs into your tools and clicks through them the way you do, so it can operate the twenty-year-old supplier portal nobody will ever build an integration for. For most businesses that matters more than anything clever the model can write.

You can also teach it by showing it. Record up to ten minutes of yourself doing a task and the Bot turns those steps into a skill you can schedule, though xAI admits the result is a draft you still have to add rules to.

It is bundled rather than sold on its own, so you need Cursor Premium Teams at $120 a seat, Cursor Ultra at $200 a month if you are buying for yourself, or SuperGrok Heavy. Mac, Windows and iPhone at launch. The eight roles it ships with tell you who it was built for: sales outbound, talent scout, paid media, expense manager, product performance, bug reproduction, account health, chief of staff.

The line in the docs worth knowing before you connect anything

All of your Bots share one cloud computer, with the same files, browser sessions and logins. Each gets its own screen, and the docs carefully call those screens separate work surfaces rather than separate security boundaries. So the Bot screening your candidates can reach whatever the Bot filing your expenses signed into, and deleting a Bot does not clear the sessions it left behind.

To be fair to xAI, the loudest criticism going around is wrong. You are not handing passwords to the model. When a login or a two-factor code comes up, the Bot pauses and gives you the screen, you type it in, and you hand control back.

Three more things matter before you connect anything real. There is no rehearsal mode, and the docs say plainly that a test run does real work on real sites and files. Approval boundaries are sentences you write rather than switches you flip, so the agent only respects fences you thought to describe in advance. And the audit view is still listed as coming, which leaves the chat transcript as your only record. None of that makes it unusable, but it does mean keeping sending, spending and publishing behind your own click for now.

Where Hermes, OpenClaw and Codex actually sit

Hermes Agent from Nous Research is open source and already does most of what people assume is unique here. It reaches you across more than twenty channels including Telegram, Slack, WhatsApp and email, schedules work with a built-in cron, spawns sub-agents, drives a browser, and writes its own skills as it goes. It is model-agnostic too, so you can point it at any provider or run something local.

OpenClaw started this category and works along the same lines, with a very large community skill library. Your hardware, your maintenance, your job to vet whatever you install.

ChatGPT Codex and ChatGPT Work sit inside OpenAI’s hosted sandbox and are strong on code and documents, but they are shaped around sessions and triggers rather than a machine that never sleeps.

So the difference is less about capability than about how much of the agent you have to run yourself. With Hermes and OpenClaw you choose the channels, the schedule, the model, the skills and the box it all lives on. Nous now offers Hermes Cloud, a hosted instance billed from your credit balance, so even hosting is no longer the real dividing line. What Grok Bot removes is the decision making around the setup itself, and you pay for that in money and in control over where your data sits.

7 Grok Bot use cases I would try

The documentation pushes you toward repeatable outcomes rather than vague responsibilities, which is good advice whichever tool you pick. These are the fifteen I would try first.

1. Chief of staff: Review email, Slack, calendar and meeting notes, then surface the few decisions that need a human.
2. Expense management: Match receipts, flag missing information and prepare the weekly reconciliation for finance.
3. Account health: Combine usage, support, billing and renewal data to flag accounts needing attention.
4. Office operations: Handle repetitive forms, invoices and onboarding steps across systems with no clean integrations.
5. Weekly reporting: Pull the same figures from several systems into a consistent management report every week.
6. Research monitoring: Follow companies, technologies or markets and report only what matches criteria set beforehand.
7. Repeated browser work: Do a workflow once while the Bot watches, correct it, save it and turn it into a routine.

What I think Grok Bot gets right

Grok Bot has not made Hermes, OpenClaw or Codex obsolete, and in several areas those tools already match or exceed its underlying capabilities. What it has done well is remove a lot of the mental setup. Instead of thinking about gateways, cron jobs, browser providers, skills, sessions, models and orchestration, you can think:

Who should own this work?

What should they be allowed to do?

When should they come back to me?

That may sound like a small interface change. For non-technical users trying to move from chatting with AI to delegating real work, I think it is the part that matters most.

An agent that arrives with a computer, a login and a calendar stops being software you use and starts being a worker you manage. That is an old skill, and judging by this week, most of us are out of practice.

Reply and tell me the first job you would hand over. I read every one.

Interested in travel or photography, read last week’s LensLetter newsletter about what’s new in Adobe Lightroom August release.

Read last week’s JustDraft about provide context to your AI agent to make personal AI Agent.

Two Quotes to Inspire

Delegation is a design decision. Anything you cannot inspect, you have already approved.

Automation copies your standards before it copies your speed, so fix the standards first.

One prompt to steal

Use this setup prompt when building your multi-bot operating structure:

I want to set up an automated multi-agent workflow for my daily operations. 

Here is a summary of my role, weekly priorities, core communication channels, and repetitive admin bottlenecks:
[Insert your daily responsibilities, tools, and main friction points]

Based on this information:
1. Recommend the ideal fleet of 3 to 4 specialized agents I should build.
2. Define the exact responsibilities and boundary for each bot.
3. Specify which bot should act as the central Chief of Staff to route my requests.
4. List the recurring routines and schedules each bot should run.