The Agents workspace

The problem

I use several AI coding tools every day, including Claude Code, Codex, Cursor, Gemini CLI, OpenCode and Hermes. Each keeps its own instructions, memory and settings. Three problems kept showing up:

• Lost context. Code and docs do not capture why a decision was made. Every new AI session starts from zero, so long tasks get rebuilt instead of resumed.
• Tool lock-in. Rules stored inside one vendor’s product are lost when the product changes or a better one appears.
• Drift. Instructions copied by hand across tools quietly fall out of sync.

My role

I created the workspace and I am its only maintainer. The repository is private, so this is a write-up of the design.

Key decisions

One source of truth. The guiding rule is that tools adapt to the workspace, never the other way round. Each tool gets a thin adapter. Any difference between a tool and the workspace is recorded and either fixed or approved on purpose.

Organize by purpose. Reusable procedures (skills), specialist behaviors (roles), tool bridges, per-tool mappings and larger workflows each have a home. A short always-on file and an index load more detail only when needed, which keeps each session’s context small.

Plan, execute, reconcile in fresh phases. For serious work, each phase talks to the next only through files on disk. Any phase can restart in a new session without relying on chat history.

Durable memory with owners. State, decisions, specs and a code map each have clear update rules. The git log is the changelog, and each commit message says why a change was made.

Rules for AI workers. Tasks go out in bounded packets. Workers cannot start other workers. The top-level agent keeps architecture and final checks.

Safety boundaries. One tool’s configuration changes per session, routine permissions are defined up front, and sensitive material lives in a separate private repository.

Outcome

Since June 2026 the workspace has grown past 1,400 commits, with adapters for six AI tools. Work now carries over between sessions and between tools. Switching to or adding a tool means writing an adapter, not starting over.

This website was built with it. Research, planning and decisions for the site are recorded in the same workspace.

What I learned

• The reasons behind a decision are worth more than the code that came from it, and they are the first thing lost.
• Short always-on rules plus on-demand detail work better than one giant instruction file.
• Giving AI workers a narrow, written task gets far better results than a vague one.
• Plain text in git outlasts any single tool.