Seven Wonders of Agentic Coding

“Man’s desire to build something that will outlast him is the true wonder behind every wonder" — Anonymous

There’s a famous scene in Fawlty Towers where Basil Fawlty’s car refuses to start. At first he employs standard practice in such a situation: turn the key, turn it again. And again. He offers some semi-optimistic motivational talk to his machine: "Come on...", "Start!", and soon escalates his managerial intensity to shouting and cursing at his metal companion. But the car simply won't start. Eventually the relationship breaks down completely as Basil counts to three, climbs out of the car, slams the door, delivers a stern lecture about unacceptable performance, and finally procures a nearby tree branch and begins beating the car with it.

Actually, it's better if you see it:

If I Just Prompt Harder...

Which brings me to AI coding agents.

Let me begin by saying: They're Absolute Magic most of the time!

There are rare occasions in development that I've hand-written more than 100 lines of code (yeah—I remember those days of hand-written code like it was only last year! 😄), and the code just worked first time. Almost always, the next step after a big code dump was a big fix-up session to iron out all of the unintentional stupid. But I do remember one time being completely shocked at two pages of untested new code (with a number of indents!) working first time. I elatedly and desperately tried to tell the nearest person what an amazingly rare and beautiful thing that really was, and, like Halley's Comet, this might not come around again for years. My then 8-year-old daughter didn't share any sense of the occasion.

But for the latest agents, that's just Tuesday. I've observed I think around November/December 2025, that coding agents had gotten noticeably better at getting large amounts of code working first time—certainly waaayy better than me at this! But there's still that one time...it just can't get it right! Prompt after prompt after prompt to "Still not working" to "Try this..." to "No I've already given you the output that shows that cannot be the problem - please remember this" to "Write up an assessment of the problem, and why you keep getting it wrong and your best assessment of the underlying root cause" to "THIS IS NOT ROCKET SCIENCE!!" to reaching for a nearby tree branch...

There are still some—increasingly rare—times when I could fix or implement something by hand quicker and better than an AI. Not many, but still some! And yet why wouldn't the agent struggle? It's surprising it doesn't struggle more often! It didn't have my tools—my browser access, my DB access, my log access. It was flying blind. I was expecting something I wouldn't expect of myself.

As an aside, I've noticed an interesting phenomenon in my psychology that I tend to develop a growing stubbornness to "manually" fix a bug myself the more an AI has failed. If I Just Prompt Harder™. Also, a little worryingly, I'm growing a reluctance to have to shift my brain from "code orchestrator" to "code writer" mode. It seems like such an inefficient context switch, when I was progressing the app so much quicker as Chief Orchestration Officer than as Code Monkey With Typewriter.

Giving Agency to Agents

And yet we know that the coding agents we use are incredibly capable. After sustaining a sufficient quota of battle scars, I started to think about better solutions. The core issue wasn't agent stupidity—it’s that the agent doesn’t natively have access to the same tools we have as developers. The all-powerful genie was squarely bottled. So I started to think about the tools I had to use to fix stuff―DB clients, the browser console, log files, and access to the rest of the development stack.

So I remember the day that I stopped mid-problem-going-nowhere (very difficult for someone who thrives on closure!) and started to research how to give the agent access to these sorts of tools. Make it their problem.

I'm not a natural salesman, so without wanting to oversell an "and I got a 10x boost in productivity", I can genuinely say I did get some profound "wow" moments along the way. The agent started to self-evaluate solutions, realize it was wrong, iterate, re-test, and keep doing all this until it could prove its solution actually worked. Yes, it costs more tokens. But “please try again” fifteen times isn’t exactly cheap either—and it’s considerably worse for the soul.

So, with an agent now armed with a suite of tools that gave it developer-level visibility to the local dev stack, results were pleasing—I would say inspiring even! Inspiring enough for me to...yeah that's right, write a blog post about it!

Renewed Sense of AI Progress

This lifted my AI coding experience from the very human-in-the-loop approach that hadn't really changed for a year, to a much more—yes—autonomous—agent experience. Fix this, here are the tools, make sure your fix works, don't come back to me till you're done. Motivation and psychology are an incredibly important part of life, and this had the strongly motivating benefit of making me feel like I was actually "keeping up with AI" - whatever that meant in my troubled mind!

And so I pledge to you a 7-part series that will shed light on what I've configured for my coding agent and what effect it has had. I call the series, based mainly on the specific number of these, the Seven Wonders of Agentic Coding. On the Seven Wonders, back to Basil:

I really should have led this article with: "I gave 7 tools to my agent and got 100x productivity". But no that didn't happen. It did, however, make a noticeable difference:

• Velocity: certainly something of a development speed increase as the agent did things that I was doing manually―it was just faster!
• Sanity: significant frustration decrease
• Focus: less involvement in the weeds as the agent did more weeding―meaning less context switching from high-level to low-level affairs
• Confidence: most satisfyingly, I'd lost the feeling that I'd fallen off the AI coding train that everyone else seemed to be riding so easily and successfully

I don't have a swarm of 100 agents. I'm not a 100x developer. But I do have a pretty powered up agent that can battle the world for me for long periods. This Is Much Better™. I mean much better than what I had before. Not better than the 100-agent swarm―I mean who wouldn't want a 100-agent swarm, right?

How much will this series move the needle for your agentic development? I'm hoping lots―it did for me! And that is my aspiration for you too―may it be supremely useful. Agentic coding eh―what a wild ride! O where will it end?!? Far from here―but exactly where...I'm not sure!