Why building at the cutting edge creates fake work, and the simple setup that replaced two weeks of custom code.

Building your own AI agents feels like the serious, innovative choice. Often it’s work a tool released next month will do for you.

We asked Magic Safari founder and CEO Philip Joubert, who ran OfferZen for about ten years, how his team handles its repetitive work…

Philip uses the image of a cycling peloton: don’t be the rider out in front taking all the wind. He sees a huge amount of fake work in trying to stay at the very forefront of AI, plus the temptation to look innovative on LinkedIn. Stay close to the cutting edge, but let someone else take the first hits.

“If you just waited a month or two, a lot of that work becomes unnecessary.”

Magic Safari has a mountain of data work: PDFs and emails of the kind normally handled by teams of data-entry staff. Philip’s brother Milan spent two weeks building custom agents for it, and they worked well. Then, on a whim, they plugged their command-line tools into Claude Code with some context and a few text files. The first attempt was about 90% as good as the custom build.

What makes the off-the-shelf setup work isn’t the agent. It’s a set of markdown files, plain text documents, holding very specific knowledge about how the business works. The agent is the general tool; your files are what make it good at your job. That knowledge is the part worth your time.

Magic Safari runs a main agent the team calls Dumbledore. They give Dumbledore the work, it splits the tasks between two other agents, nicknamed Gryffindor and Slytherin, and they report back, all through Notion. Philip says it scales beautifully and stays very simple.

To handle more work, they buy more Claude Code and Codex accounts rather than writing more software. Each account costs around R3,200 a month and comes with far more usage than that would buy on its own. As Philip puts it, they don’t care about the agents: the only question is whether they’re willing to pay for another account.

Every time Claude Code or Codex is updated, Magic Safari’s system gets better without anyone touching it. A custom build only improves when you work on it. Philip’s warning comes from watching people rush onto a new agent platform, then spend nights fixing security holes a month of waiting would have avoided. Custom agents still make sense in some cases, but for a lot of jobs, wait until someone else has solved the problem.

You get most of the result for a fraction of the effort, and your team spends its time on the knowledge that makes the agents useful instead of on the plumbing. When the tools improve, so do you.

Philip’s full keynote from The Open Letter Cape Town 2026 is available to members inside the Founder Collab, where he goes further than we could cover here:

You’ll also get access to 40+ masterclasses from SA founders and operators on sales, fundraising, UX, paid media and more inside The Founder Collab.

Get more SA tech and business news, tips and business-building workflows and subscribe to The Open Letter.