Two months with AI agents: what actually runs on its own

AI agents don't automate everything. Some of my processes run 100% autonomously, others sit at 80%. What I've learned about bookkeeping, research, and build vs. buy after two months of daily use.

There's a story going around that AI agents will soon automate everything. That you just lean back and watch them work. That's not quite true. Yet, anyway.

I've used Claude Code and Claude Cowork daily for the past couple of months. I've built agents for bookkeeping, social media monitoring, research, and content production. Some run fully autonomously. Others need me sitting there steering. Here's what I've learned about the difference.

Some processes are 100% automated. Others are 80%.

I have an agent that runs on a cron job at night. It scans X, Reddit, TikTok, and Instagram for what happened in the last 24 hours in the topics I follow. In the morning a report is waiting. If something looks relevant, I can ask the agent to dig deeper via Perplexity or other sources. That process rarely fails, because the task is simple: find, filter, summarize.

My bookkeeping agent is another story. It finds invoices, downloads them, analyzes them, and posts them. It's maybe 80% there. I still have to approve entries, fetch the occasional invoice it can't find on its own, and check that it categorized things correctly.

A study from Stanford and Carnegie Mellon found that autonomous agents succeed 32-49% less often than humans alone on complex tasks. But hybrid teams, where humans and AI work together, perform 68.7% better than pure automation. The accounting industry reports that 80-90% of routine tasks like categorization and reconciliation can be automated, while 20-45% of the work still needs human judgment.

My experience matches. The agent is a new colleague, not a replacement.

Build vs. buy is a different question now

The answer used to be: buy a SaaS product. Building it yourself was too slow and too expensive.

Not anymore. When Claude Code can build a working tool over a weekend, the math changes. According to Retool, 35% of companies have already replaced at least one SaaS product with software they built themselves. The average company uses 305 apps and pays $55.7 million a year in SaaS subscriptions, with prices rising 11.4% annually.

That doesn't mean you should build everything. Sometimes an existing product is better. If I want to auto-post to social media, I could build it myself against Facebook's API. But that takes time, and there are tools out there that cost almost nothing and are sensibly built.

My rule of thumb: build if more than 70% of what you need is unique to your business. Buy if a standard product covers it. Don't build just because you can.

Iteration beats perfection

Coding with AI agents usually starts in an idea phase. The first version is rarely the right one. You shoot from the hip, see how far you get, learn something, and start over with that knowledge.

That's different from traditional software development, where you plan for weeks and build for months. Here the cycle is hours. And the output per cycle is 10x what I used to produce. The time itself is maybe only halved, but the result is markedly better, because I can iterate faster and think bigger.

That opens up things you'd normally never attempt. Instead of optimizing what you already do, you can ask: what would the best-case scenario look like? And then build it.

The memory file is everything

The most important thing I've learned: agents have to remember what they learn. I had an agent that was good at writing blog posts and doing research. But it kept writing "billioner" instead of "milliarder" (the classic mistranslation of the English "billions") and dropped the Danish letters. After I corrected it once and saved the correction in a memory file, it never did it again.

That's the difference between an agent that repeats the same mistakes and one that gets better over time. Without the memory file you start from scratch every time. With it, you build.

Terminal work feels like the Commodore 64

There's something ironic about the most advanced AI technology we have being driven from a terminal. Plain text in, text out. It reminds me of MS-DOS and the Commodore 64 with BASIC, if anyone remembers those.

The way you talked to a computer back then looks a lot like the way we work with AI agents today. You type an instruction, the computer answers, you adjust. The interface is the same. The intelligence behind it is fundamentally different.

Where we actually are

We're not at the point where everything is automated. We're at the point where some processes run unattended and others need you sitting in as a kind of team lead for your agents.

The real gain is not the time saved but the things you would never have done manually. A monitoring system that runs every night. A bookkeeping assistant that handles 80% of the entries. An analysis of your investments built from Reddit data, market data, and volume, compiled into a report you can listen to in the car.

AI agents work. The question is whether you're willing to work with them as colleagues instead of expecting magic.