Building a Local AI Agent Loop That Stays Useful
A practical architecture for connecting research, drafting, review, and publishing without turning every task into a browser session.
The most reliable AI workflows are usually the least magical: small tools, explicit files, and a review boundary that is easy to inspect.
Start with a durable handoff
Markdown in Git gives research agents, writers, editors, and automation scripts a shared format. The important part is not the folder name; it is the contract around each handoff. A draft should say who created it, what state it is in, which sources informed it, and what still needs review.
Keep publishing boring
An agent can research, write, validate, and suggest links. Publishing should be a separate command that checks the article again, records a log entry, and creates a normal Git commit. That makes the public site a consequence of an auditable state transition rather than an accidental side effect of generation.
Prefer local interfaces
The same workflow can be used by a shell script, Codex, OpenCode, or a future local model. A CLI is the primary interface; a loopback-only JSON API is a convenience for agents that already speak HTTP. Neither needs access to a browser or a provider-specific SDK.
The result is a small system that can grow without hiding important decisions inside a single prompt or a privileged automation account.