Code is Temporary. Architecture is Executable.

Bounded Autonomy

SubscribeSign in

Code is Temporary. Architecture is Executable.

Arman Wolkensteiner-Jalili<br>Jul 09, 2026

Share

That sentence has been bouncing around my head for weeks. Let me explain why I think it’s the most important idea in AI-assisted engineering right now.<br>Every coding agent on the market works the same way. You describe what you want. The LLM figures out what to do step by step. It runs a command, checks the output, maybe fixes an error, loops. If you’re lucky, you get what you asked for. If you’re not, you get a broken test suite and a weird import you didn’t notice until the deploy failed.<br>This works fine when you’re watching. It breaks down everywhere else.<br>I’m a solution architect. I spend my days thinking about how to run systems at scale without them falling apart. And from where I sit, the current approach to AI coding tools has a fundamental problem:<br>We’re trying to audit conversations when we should be auditing structures.

You can’t audit a conversation in any meaningful way. You scroll back through a chat log and try to reconstruct what happened. Maybe it’s in a file somewhere. Maybe it isn’t. There’s no boundary on what the agent can do, no plan you signed off on before it started executing, no way to prove that something didn’t happen.<br>This is a problem if you care about compliance, which is becoming everyone’s problem. The EU AI Act starts full enforcement this August. You need traceability. You need budget limits. You need human oversight gates. A chat log doesn’t give you any of those things.<br>The Terraform Paradigm

So what does giving those things look like?<br>It looks like Terraform.<br>Terraform works because it separates planning from execution. You run terraform plan, you get a DAG — a deterministic list of every resource that will change, in dependency order, before anything actually happens. You review it. Then you run terraform apply. The plan is the contract between what you intended and what the system will do.<br>Nobody would use Terraform if it worked like an agent loop — “here’s a prompt about the infrastructure I want, go figure it out step by step, surprise me.”<br>Coding agents should work the same way. The LLM translates your intent into a structured plan — a DAG with explicit steps, dependencies, and boundaries. You review it. You approve it. Then it executes, constrained by the policies you set: which files it can touch, how many tokens it can spend, what validation gates it needs to pass.<br>The LLM generates the content. The architecture governs the execution.<br>The Shift to Bounded Autonomy

This is the shift from conversational AI to bounded autonomy.<br>The AI proposes. The constraints enforce. The result is auditable by structure, not by reading a chat log.<br>I keep coming back to that line. Code is temporary — it gets rewritten, refactored, replaced. But the structural boundaries around it — the policies, the validation gates, the execution graph — those are what keep a system from decaying into something nobody understands.<br>The volume of AI-generated code is about to overwhelm manual review. We can’t have a senior engineer read every PR written by a model that generates code faster than a human can scroll. The only answer is to make the architecture itself the gate — baked into CI/CD, enforced automatically, reviewed before execution.<br>Over the next few weeks I’ll be writing more about what this looks like in practice. How to translate something like the EU AI Act into automated engineering constraints. What a policy-driven agent loop actually looks like at the code level. Why I think the DAG will replace the chat window as the primary unit of AI interaction.<br>But for now: if you’re building with AI agents and the only thing between your production system and an off-script model is a human reading a diff — you might want a better architecture.<br>Software is temporary. The boundaries you enforce are what last.

Subscribe

Share

Discussion about this post<br>CommentsRestacks

Ready for more?

Subscribe

© 2026 Arman Wolkensteiner-Jalili · Privacy ∙ Terms ∙ Collection notice<br>Start your SubstackGet the app<br>Substack is the home for great culture

This site requires JavaScript to run correctly. Please turn on JavaScript or unblock scripts