Most teams let a coding agent run and then spend days cleaning up after it. Carlos Kelkboom, a lead architect, does close to the opposite: he does the thinking up front, lets agents build in one big go, and throws away anything he can’t immediately follow. He laid out his rules for coding agents on the Beyond Coding podcast with Patrick Akil. Below are the five rules underneath his method, the evidence for them, and what they mean for anyone using agents to build websites, tools or automations. The full step-by-step version is in our coding agent workflow guide.
Why rules for coding agents matter now
Agents write code quickly. The trouble starts when they have to change it. A 2026 research benchmark, SlopCodeBench, asked 15 coding agents to keep extending their own solutions as requirements changed, across 36 problems and 196 checkpoints. No agent fully solved a single problem end to end, and the best passed 14.8% of checkpoints. Code got more tangled in 77% of runs and more bloated in 75.5%. Simple rules for coding agents are how you avoid that slide.
The evidence behind the rules
Here’s the reality. I’ve built, scaled, and sold three companies, generated over $85 million in the last seven years in my own companies. In every one, the systems that lasted had a clear structure before anyone built on them. The research below says agents need the same thing, and Carlos’s restarts show why.
| Evidence | What it shows | Source |
|---|---|---|
| Best agent passed 14.8% of 196 checkpoints | Agents struggle to keep extending their own code | SlopCodeBench, arXiv 2603.24755 |
| Structural erosion rose in 77% of runs | Quality degrades with each round of changes | Same |
| Agent code 2.3x more verbose than 473 human repos | Agents add code rather than reshape it | Same |
| Second or third run usually lands after a restart | Restarting with the cause fixed beats patching | Carlos Kelkboom’s own experience (podcast) |
The five rules for coding agents
1. Decide the big things yourself. What the software is for, how it’s structured, where it runs, where data lives. Kelkboom calls these things you “do not want to let the agent decide by itself.”
2. Scaffold before the agent starts. He creates the folders, names them, puts a short note in each explaining its job, and writes the key files with function names and data shapes. It takes 20 minutes to an hour.
3. Keep a main agent that only talks. His first agent never edits files. It asks questions until it understands, then hands work to sub-agents, and only when it’s more than 90% confident.
4. Build in one big go, in a separate branch. Agents are “bad at making changes and good at doing it one shot,” in his words. So he gives them the largest chunk he can, in a Git worktree.
5. Throw away what you don’t understand. “If I don’t understand it, I throw it away.” He checks in after an hour or two, bins anything unrecognizable, fixes the cause and reruns.
AI coding agent best practices for agencies
These rules apply well beyond compilers. For a marketing team, the same habits keep agent-built work maintainable:
- Landing page templates: decide the page sections and data fields first, scaffold the components, then let the agent fill them.
- Reporting scripts: define the output table and data sources, and keep a planning chat that can’t touch the repo.
- Site fixes at scale: run the change in a branch and check the first few pages before letting it touch hundreds.
Vibe coding vs agentic engineering
Vibe coding is prompting until something works, without reading the code. That’s fine for an experiment you’ll throw away, and Kelkboom says so: if the goal is fun or a quick test, you’ve already won. Agentic engineering is what you do when you’ll live with the result: decisions, scaffolding, review and repeatable checks. The mistake is treating a product like an experiment.
Should you review AI generated code?
Yes, but not line by line at the end. Among his rules for coding agents, this one surprises people most: he reviews early, and for recognition. Does the code match the structure he set up? If not, it goes. He also asks agents to write less code so review stays possible, and he’s wary of unit tests alone because they’re easy to game. Mutation testing, which deliberately breaks code to see if tests notice, is a stronger check.
Want agent-built work you can trust?
We build automations and site changes with these rules: a spec and scaffold first, work in branches, early checks, and your sign-off before anything ships. No risk of a surprise release: nothing goes live without your yes. Why now: agents are cheap to run but expensive to clean up after. See the options below.
| What we do | For whom | Where |
|---|---|---|
| Agent-built automations with scaffolds, checks and human sign-off | Agencies and in-house teams | AI automation |
| Site changes at scale, built in branches and checked before release | Sites with hundreds of pages | Technical SEO |
FAQ
Are coding agents better at one shot or iterative changes?
In current benchmarks, one shot. SlopCodeBench found agents degrade as they extend their own code, with the best passing 14.8% of checkpoints.
What is scaffolding for a coding agent?
The structure you create by hand before the agent runs: folders, short notes on each folder’s job, key files, function names and data shapes.
Why shouldn’t the main agent edit files?
So it focuses on understanding the request. Kelkboom found agents jump to building; a talk-only main agent asks questions first and hands work to sub-agents.
Is vibe coding bad?
Not for experiments you’ll throw away. It’s risky for anything you have to maintain.
What is mutation testing?
A test of your tests: small bugs are planted in the code, and good tests should fail. It catches tests that pass without checking anything real.
Want coding agents that ship what you meant?
Book a call and we’ll show you where agent-built automation can save your team hours, safely.