AI Software Factory: the model never pulls the trigger
An AI-native software development factory that plans, builds, and validates software in production. Deterministic where it must be, chaotic only where it's safe.
Sometimes the best thing you can do for an AI system is take a capability away from the model.
As part of a small core AI engineering team, I've been working on exactly that: an AI-native software development factory that plans, builds, and validates software in production. The system runs. The model inside it is never the one pulling the trigger.
A model is the AI influencer who confidently gives you a different answer every time you ask the same question. Entertaining at lunch. Terrifying when they're building critical systems.
So we drew a hard line. Four layers, each exists for a reason.
Non-determinism stays inside the loop. Side effects live outside it.
Workflow Engine
The build is a DAG, not a script. Steps have real dependencies. Validation can send work backwards to be repaired. A graph lets us retry one step, mark downstream work stale, and resume from a checkpoint without redoing what already passed. Declared in YAML so the pipeline shape is reviewable, diffable, and version-controlled. The same workflow every run.
Agent Workers
Where the non-determinism lives. Workers run the LLM loop, but they only request tools, never act. The messy part can't cause side effects, and it never sees a credential.
Tool Gateway
Every request is approved or denied. Deny-by-default, path containment, audit logging. The model can ask for anything but only gets what policy allows.
Publisher
The one place irreversible actions happen: push, open a PR. It runs outside the loop, reads the finished run's checkpoint as a contract, and acts only if the run is proven complete. A re-run is a no-op, never a duplicate. All the chaos stays upstream. The moment we touch the real world, it's boring and repeatable.
One boundary out of several. More on how the factory is built coming soon.
Watch It Run
Four walkthroughs of the factory, from the DAG and its boundaries to the control tower that oversees a build.
DAG and Boundaries
Why the build is a graph, and where the hard lines between layers sit.
TUI-Based Agentic Workflow
Driving the factory from the terminal: watching agents work a run.
Control Plane Architecture
Inside the deterministic scheduler: retries, repair, checkpoints, resume.
Application Build Control Tower
The operator's view: observing and steering builds in production.
Building agent systems with boundaries?
This is the kind of system I design and ship. Let's talk about yours.