Agent workflow (how to build software with 100% autonomous coding agents)

X07 is designed around a simple reality:

An agent that writes code also needs to run checks, read diagnostics, apply patches, and prove invariants—without human intervention.

So the workflow is built into the ecosystem:

• strict output contracts (JSON everywhere),
• deterministic tools,
• stable error codes,
• canonical file layouts.

The “architecture contract” file

Every serious X07 project should include:

• AGENTS.md

This is a short, structured contract that tells the agent:

• what the system is,
• what constraints it must respect,
• what must never change,
• how to run tests and interpret failures.

Use the template in: Agentic design patterns.

The minimal agent loop

An autonomous agent should follow a loop like:

1. Read task/spec + AGENTS.md
2. Modify code only through structured patches (JSON Patch)
3. Run deterministic checks:
   • x07c lint
   • x07 test
   • optional: benchmark suites (for performance budgets)
4. If it fails:
   • parse x07diag output
   • apply a suggested quickfix (JSON Patch), or produce a new patch
5. Repeat until green

If you want a good mental model for “AI-native engineering”, see OpenAI’s Codex guide on building AI-native engineering teams.

Golden rule: keep boundaries explicit

Agent-written code is far more reliable if your system architecture makes boundaries explicit:

• pure logic separated from I/O,
• deterministic tests separated from OS worlds,
• adapters kept thin and declarative.

The “ports and adapters” (hexagonal) model is a good default because it keeps I/O at the edges and preserves a stable functional core. See Alistair Cockburn’s write-up for background.