87 Real AI Agent Tasks: What Actually Worked (and What Flopped)

One developer logged 87 ways he used an AI agent on his own computer. Here is what actually worked.

A developer using the AI agent OpenClaw kept a running list of every task he handed to it over several months, starting January 22nd. The list runs to at least 87 entries and reads less like a product demo and more like a field journal: some jobs took seconds, some took dozens of back-and-forth attempts, and several never worked at all. Taken together, it is one of the more honest looks at what a personal AI agent can and cannot do in 2025.

The author is anonymous in the post, but the setup is specific. He runs OpenClaw on a MacBook, with a Dell XPS 7590 repurposed as a network drive, a rented server in Germany from the host Hetzner, and a password vault from Bitwarden that the agent can unlock to log into services on his behalf. The agent has access to his terminal, his files, his calendar, his notes, and a managed web browser.

The wins are mostly plumbing, not magic

The most useful tasks on the list are unglamorous. The agent freed up 106.8 GB of disk space by tracking down forgotten files: 58.9 GB of an old MySQL database hiding inside the Homebrew package manager, 48 GB of iOS Simulator runtimes, 28 GB of Android developer tools, and 10 GB of screenshots. It diagnosed a slow terminal startup, found that the Node version manager was responsible for 61% of the delay, and rewrote the config to load it only when needed. Boot time dropped from 5.3 seconds to 0.78 seconds.

It set up a server from scratch: installed Docker, configured the workflow tool n8n, added an Nginx web server, wired up automatic HTTPS certificates through Cloudflare, and later patched the same server when Hetzner emailed a security warning. It remembered how to log back in weeks later, which the author calls out as the moment he was genuinely impressed.

Other quiet wins: merging PDFs, renaming downloaded files to match metadata found online, generating an invoice by duplicating a Word template, updating an Excel log, exporting to PDF, finding the client's email in Gmail, and sending the attachment. That last one was a single prompt covering five separate apps.

The failures are just as instructive

Several entries on the list end with the author giving up or stepping in manually. The agent could not log into his Spanish bank to download tax receipts because it kept getting stuck at the two-factor authentication step. He tried a competing tool, Claude Cowork, and burned through 100 tool calls without success. A grocery order at the Spanish chain Mercadona only worked after he switched to a special browser extension, because anti-bot systems blocked the other approaches.

Setting up the file-sync service Dropbox on Arch Linux defeated every model he tried, including Anthropic's Opus and the latest models from OpenAI. A daily logbook in Apple Notes worked once and never again. A "coach" routine that was supposed to check in with him morning and night worked in theory but he stopped using it after a few days.

Several other entries note that he had to step in to fix code the agent wrote, configure a service it could not reach, or restart a task that ran out of memory mid-job. The fully autonomous framing that often surrounds AI agents does not survive contact with the list.

A few patterns worth copying

The author gives the agent a dedicated password vault that only contains credentials for services it is allowed to touch. That is a sensible isolation pattern: if the agent goes wrong, the blast radius is limited to what is in that vault. He also has the agent back up its own memory and configuration to a private GitHub repository once a day at 6 PM.

He credits a principle called Gall's law: complex systems that work tend to start as simple systems that work. His early use cases are almost trivially small, things like checking disk space or opening a browser. The ambitious workflows came later, once he had a feel for what the agent was reliable at.

One practical tip from the post: when using non-Anthropic models, he tells the agent something like "try looking for other options if it doesn't work, don't give up on the first attempt." He found it improved how the model handled tool failures.

A couple of things to take with a grain of salt

The post mentions a feature where the cloud platform Vercel released a browser specifically for AI agents, and claims the author pointed his bot at a skill file to teach it. The post does not link to an announcement, and the detail has not been verified against a primary source. Treat it as a user report, not an established fact.

The author also says his Spanish internet provider blocked the finance app MoneyLover because of a court order tied to the soccer league LaLiga. That is plausible (LaLiga has obtained broad blocking orders in Spain) but the specific cause for this user's outage is his interpretation, not a confirmed fact.

Finally, the list appears to be cut off mid-sentence at entry 87, and some entries between 67 and 87 are missing or out of order. The takeaway is the texture, not the exact count.

What This Means for You

If you are thinking about giving an AI agent access to your computer, this list is a useful reality check. Three concrete things you can take from it:

• Start with a single boring task. Disk cleanup, file renaming, calendar lookups. You will learn how the agent thinks before you trust it with anything that matters.
• Isolate its credentials. Create a separate password vault or account that only holds logins for services you are willing to let the agent touch. Do not give it your main password manager.
• Expect anti-bot systems to win. Anything involving two-factor codes, CAPTCHAs, or banking logins is likely to fail. Plan for the agent to save you time on the boring 80% and leave the gnarly 20% to you.

The honest pitch from this list is not that an AI agent will run your life. It is that, used carefully, it can quietly handle a lot of the small computer chores that pile up, and occasionally surprise you by recovering 100 GB of disk you did not know you had lost.