A pattern for composing AI agent workspaces declaratively from versioned, layered sources — so the agent's operating context is reproducible, auditable, and free of drift.
If your team uses AI coding agents across multiple products and creates short-lived workspaces frequently, you eventually hit two problems at the same time:
- Drift. The transversal files that shape your agent's behavior — rules, skills, workflows — get edited inside individual workspaces. Soon you have eight slightly different versions of the same rule, and consolidating them is painful enough that nobody does it.
- Bloat. Every new workspace inherits everything, including capabilities that don't apply to the current product. The agent's context fills with irrelevant material, costing tokens and degrading focus.