What I’ve learned about building a system around coding agents

A lot of the conversation around coding agents is still about prompts.

How do I write a better prompt?
How do I give Claude more context?
How do I make the agent follow instructions?

I think there is a more useful question:

What system should exist around the agent?

I've been working on a Software Factory model with six layers:

1. Boundary
Where can the agent work? What permissions and isolation does it have?

2. Context
What repository and organizational knowledge should it receive?

3. Skills
What reusable engineering workflows should agents be able to execute?

4. Execution
How do tasks actually run in isolated environments?

5. Verification
How do we test and check what the agent produced?

6. Delivery
How does verified work reach a human-reviewed pull request?

The important part is that these layers solve different problems.

A bigger prompt doesn't solve poor isolation.

More context doesn't solve missing verification.

A smarter model doesn't automatically give you a safe delivery process.

I'm running a free live series on Maven where I walk through how to build these layers.

The first two sessions are:

The later sessions cover automated verification and background coding agents.

If you're building with coding agents on a real codebase, I'd be interested in how you're handling isolation and verification today.