From OpenClaw to Grok Bot, every AI agent still needs a job worth doing

In March 2025, an AI agent called Manus went viral. You gave it a goal and it worked through the steps on its own, and the company said two million people joined its waitlist in seven days. Within a week, TechCrunch’s tester watched it crash while ordering food, fail to book a flight and leave a game unfinished.

Eighteen months later the agents are far more capable. They run around the clock on their own cloud computers, answer you on WhatsApp and check your accounts while you sleep. But read what the people who make them and use them well keep saying, and one thing hasn’t moved at all.

So here is my argument. From OpenClaw a year ago to Grok Bot today, the agent that helps you is the one you give a specific job: what to do, when to do it, what it may not touch, and how you’ll know it worked. The models keep getting better. That part is still on you.

Eighteen months of agents

The list moves fast, so here is the short version.

• Manus (March 2025). The first one most people heard of. It has had a strange ride since: Meta bought it in December 2025, China ordered the deal withdrawn in April, and the founding team resumed independent operations in September with Manus 2.0.
• OpenClaw (early 2026). A free agent you run on your own computer and talk to through messaging apps. It started in November 2025 as a side project by Peter Steinberger, went through two other names, and passed 100,000 GitHub stars by late January. Today its repository shows about 393,000.
• Hermes Agent (February 2026). Nous Research’s open-source agent with memory, skills it writes for itself, and scheduled jobs that run without you.
• Perplexity Computer (February 2026). A cloud worker that researches, builds and runs on a schedule, paid for in credits. It disappointed a lot of early users on cost. One subscriber wrote that a single software task used $100 in about an hour and a half without the result they asked for. It is still on sale, and Perplexity’s overall revenue was reported up 50% in the month after it launched.
• Grok Bot (August 2026). xAI’s always-on agent, and the one everyone is talking about this fall. Bloomberg reported 418,000 weekly users a month after launch.
• Muse (September 2026). Meta’s personal agent, reachable in WhatsApp. Reuters reported 2.8 million downloads in two weeks.
• dots and Claude Duet (fall 2026). OpenAI began rolling out dots on September 29, and Anthropic opened a preview of Claude Duet on October 21. Both work between your conversations and check in on their own.

Most of these are still shipping. What changed is how much they can do without you. What didn’t change is what they need from you.

The makers keep telling you the same thing

Look at how the companies describe their own products. xAI: “A Bot is a worker you create inside Grok Bot for a specific job.” Manus, in its 2.0 launch: “Manus empowers you to build. You decide what’s worth building.” Microsoft says its Copilot Cowork agent “performs best when instructions are clear and specific, including details about recipients, formatting, and expected outcomes.”

The people running them day to day say it more bluntly. VantaSoft, a company that runs 13 Hermes agents for clients and sells that service, put it this way in its review: “Installing takes minutes; defining the work is the real project.” A Grok Bot user who runs eleven bots to prepare for trading credits three rules: one job per bot, a written scope for each, and a separate bot that checks the others’ work. None of them is allowed to place a trade.

You can see the difference in a single product. When DataCamp tested Perplexity Computer, it asked for a comparison of eight named coding tools, with set fields, original sources, contradictions flagged, and a memo for a stated audience. It got a usable table in eight minutes for about $2.26 in credits, plus a half hour of checking and editing. The same review warns that costs are hard to predict and that open-ended software projects are a much worse match. A scoped research job and an open-ended build are very different requests, even in the same tool.

A clear job isn’t enough on its own

Here is the honest complication. Clear instructions don’t guarantee anything.

In February, Summer Yue, a director at Meta, described how her OpenClaw agent started deleting her inbox in bulk. She had told it to ask before acting and then told it to stop. By her account, the instruction got lost as the agent worked through a much bigger inbox than the one she had tested on. In October, security researchers at Salt Labs showed that a single malicious email could hijack Manus while it was doing a perfectly reasonable job: checking an inbox. The flaw was fixed before they published. OpenClaw’s own founder warns that this kind of attack, prompt injection, “is still an industry-wide unsolved problem.”

So the job description has to include the fence, not just the task. Grok Bot’s security documentation recommends spelling out which changes a bot may make and where it has to stop. That is the same intentionality, applied to what the agent is not allowed to do.

What a good job looks like

The examples that work have the same shape. A trigger, a source to check, a result you can confirm quickly, and a line it can’t cross. Here is what that could look like for a small business. These are examples, not tested setups:

• A plumber: “Every evening, list tomorrow’s jobs with the address, the customer’s note and whether parts were confirmed. Flag any job with no confirmed parts. Don’t message customers.”
• A restaurant: “Every Monday, compare last week’s supplier invoices with the order sheet and list anything we were charged for but didn’t order. Don’t contact suppliers.”
• A real estate agent: “When a new listing goes live in my three towns under $800,000, send me the address, the price and the listing agent. Read only.”

And the jobs to keep for yourself: deciding what matters, what to promise a customer, and anything with legal or money consequences. Let the agent prepare those. Don’t let it decide them.

To be fair about the limits: no study has compared the same agent given a vague goal and given a specific job. That part is my judgment, built on what the makers say, what experienced users report, and how the failures above happened. Most of the user stories here are self-reported, and several products are still in preview.

The agents will keep changing names. The job description is still yours to write.