"AI infrastructure" usually means GPUs, model serving and vector databases. For teams putting coding agents into production, there is a second stack, and most of it did not exist two years ago. This post maps it, layer by layer, so you can see what you already have and which layer is usually missing.
The stack, from the model down to Git
1. Models and gateways. Where the intelligence comes from: hosted providers, cloud platforms like Bedrock, or local models, often behind a gateway that routes requests and tracks spend.
2. Harnesses. The agent itself: Claude Code, Codex, opencode, Cursor's agents, or nool code. A harness turns a model into something that reads files, runs commands and edits code.
3. Compute and sandboxes. Where agents run. A laptop, a CI runner, or cloud sandboxes such as Daytona or Modal when you want dozens of agents at once. Nool's fleet pools route work across them.
4. Coordination and control. Which agent may change which code, which rules every change must pass, and who has to approve. This is the layer that decides what lands.
5. Source control and forges. Where the code lives: a git server, GitHub or GitLab, Cursor's Origin, or at the largest scale something like Anthropic's Scion or OpenAI's Pando.
6. CI and review. Builds, tests and human review after a change is pushed.
The missing layer
Layers 1, 2, 3, 5 and 6 are crowded with good products. Layer 4 is usually empty. Teams fill it with conventions: one agent per repository, a human merging by hand, a long AGENTS.md asking agents to be careful.
Conventions do not hold under load. Two agents on two branches can rewrite the same function, and git only notices when they meet in CI. An agent can wander into authentication code while fixing billing. A rule written in a markdown file is a request, not a control.
What the coordination and control layer does
Nool is version control designed for agents working at once, and it sits in layer 4, on top of the Git you already have:
- Concurrency. Agents claim the code they will change before they start. A second agent reaching for the same code is refused before it writes, so collisions never reach CI.
- Safety. Tests, configurable YAML rules and hard WASM rules run before a change lands. A change that breaks them is refused, the same way every time.
- Scale. More agents in parallel without the cleanup. In our coordination benchmark with 35 agents on one codebase, 98 to 100% of tasks landed with Nool, against 33 to 62% on git without coordination.
- Evidence. Every landed change carries its intent, tests and approvals in a signed record.
It runs locally and checks a signed licence offline. It works with every harness in layer 2, and it mirrors every landed change to the forge in layer 5, so nothing above or below it has to change.
How to tell whether you need it
You need layer 4 if any of these is true:
- Two of your agents have ever edited the same function in the same week.
- You cannot say, for a given change, which task asked for it and which rules it passed.
- Your review queue grows faster than your team.
- You want to add more agents, and the thing stopping you is cleanup, not cost.
Read more about agentic development infrastructure, compare the agentic version control options, or put Nool on your repository.