Modern software teams increasingly rely on AI coding agents to accelerate development across complex, multi-repository systems. However, agents operating within a single repository lack the cross-boundary context required to reason about the full system — resulting in shallow suggestions, missed integration constraints, and repeated manual correction by engineers. This working paper presents a lightweight architectural pattern — the Virtual Workspace Meta-Repo — that addresses this limitation without requiring a monorepo migration. The meta-repo acts as a centralised scaffolding layer that clones all isolated project repositories into a single logical directory tree, giving AI agents unified access to frontend, backend, infrastructure, and Shared library code stays within a single session context. The underlying repositories remain entirely unchanged: their Git histories, CI/CD pipelines, and deployment lifecycles are fully preserved. The pattern is supported by four mechanisms detailed in this paper: a mandatory README hierarchy for agent navigation across polyglot codebases; a centralised, agent-agnostic SKILL system that separates custom AI instructions from proprietary tool formats; standalone external connectors that proxy live system queries — such as logs, tickets, and databases — without revealing credentials to the prompt; and an optional team-scoped MCP (Model Context Protocol) server that presents high-frequency tools as always-available agent capabilities. The combination of connectors and a team MCP has proven to be more effective in practice than either approach alone. The pattern is validated through production at enterprise scale. This paper details the full implementation reference, architectural trade-offs, known risks with mitigations, and a publicly accessible reference implementation.
Ivan Baha (Fri,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: