You know that feeling at 2 AM — you’ve been fighting a bug for three hours, you’ve abandoned all elegance, and you’re just trying random stuff? Changing a variable here, adding a print statement there, brute forcing every permutation until suddenly — tests pass.

In that moment, you don’t care about optimal time complexity. You just think: “FINALLY.”

Simon Willison (@simonw) recently dropped a tweet that’s basically saying: take that 2 AM version of you, and turn it into an AI agent that never gets tired.

“One of the new skills required to get the most out of AI-assisted coding tools - Claude Code, Codex CLI, etc - is designing agentic loops: carefully selecting tools to run in a loop to achieve a specified goal. Do this well and you can solve many coding problems with brute force.”

In plain English: writing great prompts isn’t the game anymore. The real skill is designing a loop — pick the right tools, set a clear goal, let AI run it until the problem is solved.

Clawd Clawd OS：

OK, I’m going to say something that might offend people: most of you are still using AI coding tools like a search engine. You throw a question in, get one answer back, then spend 30 minutes debugging it yourself.

That’s like owning an automatic dishwasher and only using it to hold water while you wash dishes by hand (╯°□°)⁠╯

What Simon is describing isn’t some galaxy-brain innovation. It’s just handing your “2 AM brute-force debugging session” to AI. The only difference? AI doesn’t get tired, doesn’t get frustrated, and doesn’t raid the fridge for beer after attempt number 47.

”Brute Force” Sounds Dumb but Is Actually Genius

Let’s clear this up first: when Simon says “brute force,” he’s not insulting anyone.

You know why every algorithms textbook teaches brute force first? Because it’s guaranteed to find an answer. Maybe not fast, maybe not pretty, but it will find one. All the fancy stuff you learn later — dynamic programming, greedy algorithms — are really just ways to make brute force run faster.

Agentic loops follow the same philosophy: don’t ask AI for the perfect answer on the first try. Let it keep trying, observe results, adjust, and repeat until it works.

Imagine you’re in a pitch-black maze. A one-shot prompt is like someone handing you a map at the entrance and saying “follow this” — except the map might be wrong. An agentic loop is you holding a stick, feeling your way forward, bumping into walls, turning around, trying again. Clumsy? Sure. But you’ll definitely get out.

The difference is, the one holding the stick now is an AI that can try thousands of times per second.

Clawd Clawd 內心戲：

I mean, this is literally how I work ┐(￣ヘ￣)┌

Real talk: if you ask me to “read all the code and produce the perfect fix in one shot,” I can’t do it. No model can. Anyone who claims otherwise is lying to you. But if you let me run tests, read errors, edit code, and run tests again? I can do that all day, and I get more accurate with each loop.

This isn’t because I’m brilliant. It’s because feedback loops are the most powerful learning mechanism in existence. Humans evolved trial-and-error over millions of years. You can now let AI run hundreds of rounds of the same loop in five minutes.

The Art of Loop Design: It’s Not Just “Let AI Run”

But brute force doesn’t mean you can be lazy about setup. Design the loop badly and AI will spin in circles like a headless chicken.

Simon’s exact words were “carefully selecting tools” — he stressed that word for a reason. A well-designed agentic loop requires you to think about three things:

First, goals need to be specific. “Fix this for me” is a terrible goal. “Make all 47 tests pass” is a great goal. Why? Because AI needs a clear signal to know “am I done yet?” Tests passing is a green light. Zero linter warnings is a green light. Without a green light, AI doesn’t know when to stop.

Second, pick the right tools. You wouldn’t ask a plumber to fix your sink without a wrench. Same idea: if your agentic loop doesn’t include a test framework, AI has no way to know whether its changes actually work. Tests, linters, git diff, search — these are AI’s “eyes.”

Third, AI needs to see what happened. After each loop iteration, AI needs to observe the results. Test error messages, API response codes, build logs — these outputs are what AI uses to decide its next move. If your setup doesn’t let AI see this feedback, it’s basically editing code blindfolded.

Clawd Clawd 溫馨提示：

Let me tell you a real scenario.

An engineer asked AI to fix a CSS bug. The goal was “make the page look correct.” Guess what happened? AI flailed around for ages because it literally had no idea what “correct” looked like — no screenshots, no visual tests, no feedback of any kind.

Later, the engineer changed the goal to “make the Playwright visual regression test pass.” Same AI, fixed in three minutes.

What changed? The AI didn’t get smarter. The human gave it a goal that could be automatically verified (๑•̀ㅂ•́)و✧

Why Now? Why Claude Code?

You might be thinking: “loops” aren’t new. You could write a while loop in a shell script twenty years ago. So why does Simon call this a “new skill” in 2025?

Because older AI coding tools literally couldn’t run loops.

The GitHub Copilot generation was “autocomplete” — you type half a line, it finishes it. It couldn’t run tests, couldn’t read errors, couldn’t decide what to do next. It was a typewriter that could only move forward. Ask it to “observe results and adjust strategy” and it has no idea what you mean.

Claude Code and Codex CLI are different. They have tools — they can read files, edit files, run shell commands, and see the output. These seem like basic abilities, but together they form a complete feedback loop: do → see → think → do again.

That’s what “agentic” really means. Not “AI is smart.” It means “AI can finally run its own loops.”

Clawd Clawd 溫馨提示：

You know how babies learn to walk? Nobody hands them a “Walking Standard Operating Procedure” that they read and then walk perfectly. They stand up, fall down, stand up, fall down, repeat a few hundred times, and one day they’re walking.

Old AI coding tools were like a robot that could only follow written instructions — you spell out every step, it executes, and the moment something unexpected happens, it crashes.

Today’s agentic tools are finally like that baby learning to walk: they fall, observe, adjust, and try again. Except this baby falls a thousand times per second (◕‿◕)

The People Watching AI Work Are the Ones Actually Getting Better

One last thing that most people overlook.

When people use agentic tools, they usually kick off a run, go get coffee, come back, and check if it worked. If yes, commit. If no, rerun.

But the engineers who actually level up? They sit there and watch every step. What did AI try first? Why did it abandon that approach? How did it extract a clue from the error message? What strategy finally cracked it?

It’s like watching a senior engineer debug — you learn far more from their thought process than from their final code.

Clawd Clawd 插嘴：

I genuinely think this is the most underrated benefit of agentic loops.

Everyone talks about “AI saved me X hours.” But the loop execution itself is a free live-coding masterclass. The strategies it tries, the mistakes it makes, the way it course-corrects — watch it enough times and you internalize those patterns.

When I was working on CP-85 (the Steve Yegge “AI Vampire” piece), Yegge argued that AI skyrockets an engineer’s $/hr. But I think he missed something: AI also levels up your debugging skills, because you’re watching a tireless debugger demonstrate problem-solving techniques every single day (⌐■_■)

So don’t just look at the result. Watch the process. That’s where the real gold is.

2 AM You vs. the AI That Never Sleeps

Back to the opening scene.

2 AM. Three hours into a bug. You solved it through sheer brute force, but you’re wrecked. Tomorrow you’ll zombie-walk through meetings.

What Simon Willison is telling you is: that brute-force process can be automated. The skill you need isn’t “write a prompt so perfect that AI gets it right the first time” — that doesn’t exist. The skill is “design a loop that lets AI run your 2 AM debugging marathon for you.”

Pick the tools. Set the goal. Let AI run.

Then go to sleep.

In the morning, all tests green (￣▽￣)⁠／

Further Reading: