Context fragmentation isolates relevant information across projects or services, so an AI agent can make a locally correct change that is systemically wrong. Each repository may contain useful instructions, source, and tests while omitting the cross-service relationships, deprecated dependencies, and operational constraints needed for safe decisions. Too little context creates a vacuum; indiscriminate access creates a flood that wastes tokens and iterations. section 10.0 section 10.1
Reduce fragmentation with deliberate system-level context:
- Assemble the relationships among repositories and services, rather than relying on project-local AGENTS.md files alone. section 10.1
- Curate a small, task-specific set of tools before generation instead of exposing the entire tool catalog. section 10.1
- Treat repository boundaries as an agent-experience decision; co-locating related services can make their relationships visible by default. section 10.1
A valid change against the wrong system
An agent working in payment-service sees its local instructions, tests, and source, but not that shared-auth was deprecated after a security incident or that neighboring services impose additional constraints. It generates a clean authentication flow using the deprecated pattern. The code passes local expectations while violating system-wide knowledge. section 10.0 section 10.1
~/projects/ ├── app-main/ # Main development, current feature ├── app-auth-refactor/ # Auth refactoring (agent working) ├── app-api-tests/ # API test improvements (agent working) └── app-docs-update/ # Documentation (waiting for review)
Parallel Git worktrees demonstrate another form of fragmentation: project state, failures, prompts, and agent knowledge split across directories and sessions, increasing coordination and merge overhead. section 5.6.1
