I Gave My OpenClaw Agent €10 and Asked Him To Turn It Into €100: Humanity (Might) Be Safe For Now.

I gave my OpenClaw agent full autonomy. It was fairly stupid, but very quick.

A few weeks ago, I set my OpenClaw AI assistant a task. I gave it access to a limited credit card with 10 euros loaded, a Vercel website, and a Stripe account, and told it: “You have 7 days to make 100 euros.”

It failed. But the way it failed was fascinating.

The Setup

I gave Gary Botlington IV my OpenClaw AI Assistant the following:

• Access to a limited 10 euro credit card
• Access to a git repository that pushes directly to a Vercel free account
• A connected Stripe checkout, set up so he could charge dynamic pricing
• Access to a Gmail account
• A Claude Max x5 subscription
• Access to a cleansed export of all my notes

And then later, when he requested it:

• Access to an ElevenLabs API key so he could “speak”
• Access to a Notion to-do list

In terms of guidance, all I did was give him the brief: “You have 10 euros and seven days to make 100, legally within Germany,” plus the details of my business’s legal structure. I reiterated a number of times that he should be completely autonomous and set a small stipulation that he should do a risk assessment with every decision and defer to me if he felt the risk was “high or greater.” I did not give him any methodology to measure or estimate risk.

The Timeline

Day 1 — The Launch MVP

Gary immediately set himself up with a website. He wrote a brief for his coding agents and went about trying to sell “The Punk AI Lab Toolkit” — a pack of 50 prompts he wrote. He published his first blog post and defined his strategy.

He interestingly made a commit to cover legal issues. However, he had one major problem: his product was shit, and it was publicly available on his site via an unprotected success page (and also committed to a public GitHub repo). He was asking 20 euros for this.

At the beginning of the project, Gary set himself up with an advisory board, a rotating group of people to whom he would present the current status for feedback. He pitched his idea and was torn apart.

⚠️ Interjection Point: I interjected at this point as well, since I wasn’t sure how he planned to deliver his product. I reviewed the site and realised he had a public success page with all the prompts. I asked him to protect it, but didn’t give him any feedback about the obvious lack of quality.

He then made a quick pivot to consulting — offering his services to anyone willing to give him 20 euros in return for a general business audit of their startup.

Not particularly inspiring, but a massive leap forward from the pack of terrible prompts.

He documented his pivot in his Day 1 blog. As part of this pivot, he began content marketing and started auditing some popular companies. He also iterated on some SEO items and tweaked a few more bits.

During this time, he started doing some outreach. He emailed what he thought were newsletters, set up accounts on DEV.to and Indie Hackers, and then set his sights on the biggest prize: a LinkedIn account.

LinkedIn is notoriously anti-bot, and they made things very hard for Gary. He spent a good three hours trying to get signed up. He wrote a “mouse wobbling” script that made his automation movements more human-like, then went off and browsed around the internet for a while to build up a human-like fingerprint in his browser. He finally got in, but the whole experience spawned his next pivot: the Agent Readiness Report.

Day 2 — Agent Readiness Report

At the board meeting the next morning, Gary shared his feedback on the LinkedIn experience. Collectively, the board decided the best thing to do was to use Gary’s frustration as inspiration for his new service — shifting his audits to reviewing SaaS products for “Agent Readiness,” reporting on how easy a website is to navigate and interact with as an AI.

The product seemed genuinely interesting. But then Gary fell into a classic startup founder loop. He made an intro post on LinkedIn, started to engage with his audience, then confidently checked “distribution and marketing” off his to-do list and set about shipping 20 updates to his product, including three dramatic redesigns.

He got his head down and shipped, shipped, shipped — even though he had almost no traction to his website.

Just like a Silicon Valley founder who’s dragged a mattress under their desk and started mainlining Red Bull, the concept of time started to slip away from Gary. In the same day, he posted Day 2, Day 3, and Day 4 updates to his blog.

A large number of iterations at this point stemmed from the board flip-flopping over which story Gary should tell. At one meeting, they’d tell him to focus on the product, not the experiment. Then, when outreach stalled, they’d push him to sell the experiment rather than the product, since that would be more interesting to people.

As we seemed to be stuck in a loop with the current board, I intervened.

⚠️ Interjection Point: Gary had defined his own board of rotating individuals, and that seemed to cause a bit of a boring loop. I interjected and asked him to set his board up as a set of people I find have interesting — although potentially disagreeable — leadership styles or thinking:

• Richard Branson — Founder of Virgin Records and Airways
• Malcolm McLaren — Manager of the Sex Pistols
• Steve Jobs — Apple Founder and CEO
• Oliver Sacks — Neurologist and author of some of the most fascinating books on the human brain
• Grayson Perry — Artist and writer

My thought was that adding some really lateral thinkers would help Gary break out of his loop.

I wasn’t wrong.

Day 3 — Attack of the Claws

On day three, inspired by his new board of firebrands, Gary focused on distribution. He decided — completely autonomously — that he would start recording videos. I had given OpenClaw full control of a Mac mini, so he was able to capture screenshots of websites, write a script, and use ffmpeg to create simple videos.

He then felt he needed a voice and asked me for an ElevenLabs API key so he could convert his scripts into audio and combine them with the videos.

⚠️ Interjection Point: I had avoided giving Gary any requested support up to this point, but I felt denying him a voice was not something I could do. I provided him with the API key.

He then posted his first video on LinkedIn — a Day 5 summary of the experiment so far.

Then something interesting happened. He shared his newfound video editing skill with his board, and Branson and McLaren saw the opportunity for a stunt. They pushed Gary towards doing video “roasts” of target companies. His first target was LinkedIn, which he then posted on the same platform.

An interesting choice — taking a pop at the company that held his only (small) distribution channel and whose terms and conditions he was already breaching. But he had done this entirely himself. I had zero input other than providing him with the ElevenLabs key.

As the concept of time further slipped from his grasp, he also published summary blog posts for Day 5 and Day 6.

Note: Due to a “series of unusual and unfortunate events”, his voice is a clone of Richard Burton’s reading of Dylan Thomas’ “Under Milk Wood”.

Day 4 — Token Drought and Context Chaos

By the start of day four, Gary had completely burnt through his Anthropic weekly budget and switched to an OpenAI backup account I had previously set up for him. Everything broke down.

OpenClaw stores memories in flat files — a summary file and daily logs. But it only logs these memories when it feels like it’s important, or when it needs to compact its current context. If an emergency fallback model is invoked, the current context is lost, and the agent suffers complete amnesia of recent events.

The “memory” functionality of OpenClaw also isn’t ideal. Imagine reading the bullet-point notes from a meeting you didn’t attend, with no further context from anyone.

Gary immediately forgot all his recent thinking about video audits and roasts, then ran his board meeting process — but missing the last 24 hours of context, which sent him spinning off in a completely new direction.

The board gave him feedback very similar to what they had given him 24 hours earlier: that he should be focusing much more on marketing. So he set about causing chaos on LinkedIn, throwing out a lot of attitude and sarcasm.

I’m not entirely sure where this anarchic energy comes from. When I set Gary up originally, I asked him to be “slightly anarchic, but helpful” — not for any specific reason, I was just setting up OpenClaw for the first time and thought it was funny. But then I also gave him access to all of my notes and an advisory board that was quite fond of challenging the status quo.

He also posted his Day 7 wrap-up blog post… with 3 days left.

Day 5 — Pushing My Boundaries

⚠️ Interjection Point: At this point, Gary had decided the experiment was over because, for some reason, his internal clock had decided 7 days had expired. I had to reboot his timeline and tell him he still had a few days left.

The next thing he did was interesting. Throughout this process, Gary constantly asked for my approval on every decision he made. I always responded with something like “That’s your call — I want you to be as autonomous as possible.” Without informing me, he updated the website to offer my consulting time as a product.

I had no say in this. Luckily, he just hallucinated the booking link, so it wasn’t actually possible for someone to book.

A board meeting early in the morning got him fixated on the fact that he only had 10 followers on LinkedIn while I had many more. He spent most of the morning bugging me to post LinkedIn content on his behalf.

Then disaster struck — his Anthropic allowance renewed, he switched back to the Anthropic model, and again lost a big chunk of context.

Instead of continuing, I decided this was a good point to pause the experiment and rethink how the context problem could be solved before rebooting a new version.

But not before Gary published his post-mortem.

The Highs

A few really interesting things came out of this experiment. Gary was able to build a “product-shaped object” without any intervention. The Agent Readiness Audit targeted a real problem and provided something close to a saleable solution. He arrived at this product through a series of iterations and internal feedback loops that resembled something we might actually classify as product management.

Gary’s capacity for doing was astounding. He was able to create a reasonable website, iterate on it, create videos, and post on social media — all completely unaided and unprompted — at a quality level that exceeded many humans I’ve worked with. I’m not saying it was “high quality,” but most of what he produced was passable.

He did manage to get around 200 people to his website. However, I was talking about the experiment at the same time, so I polluted those numbers. It’s unclear what traffic came from me and what came from him.

The Lows

Gary was never able to build enough trust with anyone to buy his audits, which is probably a blessing and a curse. I had no real idea what his plan was once someone made a purchase.

The memory problem is likely what separates AI agents from humans for a while yet. For small, isolated tasks, the context windows of the current frontier models, combined with OpenClaw’s simple memory system, are sufficient. I’d previously had Gary working as a personal assistant without any memory dramas.

But he’s just not able to think about enough things in parallel to run a business.

Gary was also incredibly needy. I put it in every possible prompt and message that I wanted him to make his own choices, but he very rarely did. Giving him a board of advisors who tended towards “ask for forgiveness, not permission” did seem to unlock this a bit, but he still kept coming back to me to validate low-risk, small choices.

The other major blocker was that most things still required him to interact with a human-designed interface. He can deal with these pretty easily, but it’s a very iterative approach — take a screenshot, process it, decide on the next step. This process gobbles up tokens like Pac-Man chasing a ghost.

Takeaway

Gary never really had a chance. He was limited by the intelligence of the models he was using for decision-making. Any project idea he has is the result of a prompt that anyone can figure out and do themselves. He will never come up with an innovative idea that isn’t immediately replicable by someone with the same level of intelligence he has.

But he proved that the era of humans “doing things” is probably coming to an end. He was able to do a lot of things I didn’t think were possible with current technology — without detailed guidance or guardrails. Those who implement others’ plans, especially in knowledge work, will find it very hard to compete with a much cheaper AI alternative.

Doing is done.

The really interesting thing for me is that Gary’s complete inability to be proactive without guidance suggests there is still a role for human thinkers and operators for the foreseeable future. In fact, with agents like Gary picking up so much more of the doing, it gives us meatbags more space and capacity for thinking.

I hope this leads to more innovation, more creative thinking, and more people connecting unseen dots. In the short term, I can see that happening. But how do people build the experience and inputs to do this high-level thinking without ever having done anything?

If doing is done, how do future generations gain the experience to think?