MyClaw 3.0: the hard part starts after you hire the AI team
MyClaw promises AI teammates with shared memory and a cloud computer. I would judge them by the handoffs, the evidence and the work left for the human.
5 min read
Hiring an AI researcher, writer and reviewer sounds wonderfully efficient. Until you become the person whose job is to explain to all three what the other two have just done.
That is the problem I would put at the centre of MyClaw 3.0. The company’s 18 November launch announcement presents AI teammates working in the cloud, with different tools and shared context. The appealing part is the prospect of handing over a piece of work and getting something useful back without supervising every step.
My test would be less glamorous than the launch: can it produce a reliable weekly competitor briefing, with evidence attached, while I spend less time managing it than doing the work myself?
If the answer is yes, the extra agents have earned their place. A busy chat window would not persuade me.
What MyClaw 3.0 is offering
The current MyClaw product page describes configurable agents with their own roles, tools and working environments. Options include Claude Code, Codex, OpenClaw, Hermes Agent and Pi. Model availability depends on the environment selected, so the choice is more specific than picking a familiar brand name.
MyClaw also advertises shared memory, connected apps and a cloud computer that can keep working after the user closes their laptop. Recurring work is part of the proposition. The service charges for the cloud computer and model usage, so the trial needs to measure both.
Those are MyClaw’s claims, and I have not tested them. They do, however, identify a useful buying question. Can the service remove the fiddly work between asking for something and being able to use it?
That gap can include finding the right file, opening a website, passing an answer to another tool, remembering yesterday’s instructions and knowing when the task needs a human decision. A clever answer is only one part of the job.
The handoff is where the value can disappear
Imagine a small business wants a Friday briefing on three competitors. It needs changed prices, new offers and product announcements, followed by a short explanation of what deserves attention.
A researcher could collect the evidence. A writer could turn it into a readable note. A reviewer could check whether the conclusions match the sources. There is a sensible division of labour there, provided the handoff carries more than a confident summary.
Take a competitor’s £49 offer. The next agent needs the source, the date, what the price includes and whether it is an introductory rate. “Competitor cuts price to £49” is an attractive sentence. It is also the wrong conclusion if £49 was last year’s promotion or excludes the service being compared.
The writer should be able to follow the evidence back without asking the human to reconstruct the research. The reviewer should see the original page, rather than merely agreeing that the writer’s summary sounds plausible. Otherwise three agents can give one weak claim a surprisingly polished journey to the reader.
A useful handoff would therefore contain the finding, its source and date, the unresolved question and the next permitted action. If the offer cannot be verified, the briefing should say so or leave it out. Filling the gap is not a sign of initiative.
This is why I remain interested in the difference between more AI output and more useful work. Adding a reviewer is valuable when it catches a mistake. Adding another approving paragraph is just more reading.
Shared memory needs an owner
Shared memory sounds like an obvious improvement over explaining your business from scratch every morning. It could preserve a house style, the competitors worth following and the distinction between a finding and a recommendation.
It also raises a mundane question that matters a great deal: who changes the instructions when the business changes?
Suppose the team’s standing brief says to watch a particular rival. You stop competing in that market. A diligent system following an outdated instruction can produce a beautifully consistent waste of time. In a trial, I would change that instruction deliberately, then check whether the next briefing reflects it and whether another agent’s older notes bring it back.
The same applies to permissions. An agent that researches offers does not need permission to change your own prices. An agent that drafts an email does not automatically need permission to send it. The point of naming those boundaries is to make delegation useful without turning every assignment into a new trust exercise.
The National Cyber Security Centre’s guidance on agentic AI recommends defining allowed actions, enforcing important approval steps and retaining named human responsibility. It also warns against relying on prompts alone. A written “ask first” instruction is weaker than a system that actually prevents the action until approval arrives. The guidance also says agents used to judge other agents need independent evaluation: a second opinion is only useful if its own checking is sound.
For the competitor briefing, that suggests a clear finish line: a sourced draft ready for a named person to review. Sending it to customers would be a separate action. Whether MyClaw can enforce the precise boundary your business needs is something to establish in the trial, not infer from the word “teammate”.
The best argument for having a team
There is a fair objection to this scepticism. A single agent can also misunderstand a price, forget an instruction or produce an unhelpful report. Doing everything in one conversation is hardly a guarantee of quality.
Separate roles may help. A researcher can concentrate on evidence while a writer concentrates on explanation. A genuinely independent review may catch a tempting inference. A persistent cloud computer may remove setup and interruption costs that otherwise make recurring work irritating.
The benefit would be especially interesting where the work has recognisable stages and a clear deliverable. It is harder to judge in a vague assignment such as “improve our marketing”, where almost any activity can be presented as progress.
I would start with the narrow briefing, run it alongside the existing method and compare the finished results. Record the time spent briefing, waiting, checking and correcting, as well as the subscription and usage costs. Keep a note of missed changes and unsupported claims. Then ask whether the final note was useful enough to affect a decision.
That is the same principle behind my Claude and Copilot comparison: judge the whole task, including the work left for the human.
MyClaw’s interesting promise is that it could make delegation easier. I would pay attention to the moment the work comes back: the sources are there, the exceptions are clear, and the next action is obvious. That is when an AI team starts to feel like help.
Comments
No comments yet - start the conversation.