The hard part isn't the agent.
It's
everything around it.
That surrounding layer is called the harness. It decides whether an agent is any use on your codebase — and unlike the model, it's the part you're left to build. Neroli ships it instead.
Neroli is currently in beta on macOS.
- Isolation, environment, checks — handled
- Keep Claude Code and Codex as they are
- Native Mac app, local or cloud
In plain terms
What a coding agent harness is.
The model writes the code. The harness decides what it can see, what it can run, where it runs, and how you know it worked. Same model, different harness, different results.
The model is the part you can't change much. The harness is the part you can, and it's where most of the difference comes from.
The catch
Half of it comes in the box.
Claude Code and Codex are already harnesses, and good ones. That's the inner harness. The outer half depends on your repo and your definition of done, so nobody ships it for you.
| Inner harness | Outer harness | |
|---|---|---|
| Who builds it | Your agent vendor | You, usually |
| What it is | Claude Code, Codex | AGENTS.md, setup scripts, worktrees, tests, review |
| What it does | Runs the loop, calls the tools, manages context | Isolates the work, prepares the environment, proves the result |
| When it's missing | Wrong tool call, lost context | Agents overwriting each other, no way to run the tests, changes you cannot verify |
| Where Neroli sits | Untouched — keep the tools you already use | This is what Neroli is |
What you get
Six things you would otherwise build yourself.
Most teams build the first two by hand, then meet the rest the day a second agent starts running.
- 01
One task, one workspace
Two agents in the same folder overwrite each other. Each task needs its own branch, code, and services.
In Neroli: Neroli makes one per session. You never set it up.
- 02
Say the rules once
Agents guess when they don't know your setup steps and commands. Write them down and the guessing stops.
In Neroli: Workspace Commands live in the repo, so every session starts the same way.
- 03
Local or cloud, your call
Some work belongs on your Mac. Some should keep running after you shut the lid.
In Neroli: Move a session either way, whenever the task calls for it.
- 04
Let it check its own work
An agent that can't run your tests is guessing. One that can fixes its mistakes before you look.
In Neroli: Tests, builds, a browser, and the iOS Simulator attach to the session.
- 05
See proof, not a summary
“I fixed it” is not evidence. You need the diff and the results together.
In Neroli: Every session shows its changes next to what the checks returned.
- 06
Make it repeatable
The same job comes back every week: the release check, the dependency bump, the triage pass. Doing it by hand each time is the tax.
In Neroli: Save it as an automation and run it on a schedule, on a trigger, or on demand.
Questions
What developers ask.
What is a coding agent harness?
Everything around the model. The model writes code. The harness decides what it can see, what it can run, where it runs, and how you know it worked.
Is Claude Code a harness?
Yes. Claude Code and Codex are both harnesses around their models. They are the half that comes in the box — the inner harness.
So what is the outer harness?
The half that doesn't: isolating each task, preparing the environment, choosing where work runs, and proving the result. It depends on your repo, so no vendor ships it. That's what Neroli is.
Does the harness matter more than the model?
Often. Whether an agent can run your tests, see what broke, and stay out of other work's way has nothing to do with which model you picked. Those are all harness properties.
Do I need a tool for this?
Not at first. A CLAUDE.md file, a few scripts, and manual worktrees work fine. It stops being free around the second or third agent.
Does Neroli replace Claude Code or Codex?
No. You keep the tools and accounts you already use, running through their own interfaces — no wrapper, no chat layer in between. Neroli builds the layer around them, and is not affiliated with Anthropic or OpenAI.
Field guide
Build the outer harness around a real Xcode app.
Follow a task from its Git worktree through isolated builds, app identity, verification evidence, pull request, and cleanup.
Read the Xcode software factory guideNeroli beta
The advanced setup, without the setup.
Tell us how you work today and which provider you want to run first.
Request access