AI Coding Agents Meet the Limits of Polyrepo Development
AI coding agents are software development automation tools that read, modify, and generate code autonomously, but their impact is constrained when they only see a single repository and cannot retain cross-project context or past work. In many teams, agents excel at local refactors yet stall when a change spans multiple services, shared libraries, or APIs maintained by different groups. They also forget previous sessions, so each task starts with a partial view of the system. This gap becomes obvious in polyrepo setups, where hundreds of isolated repositories hide dependency graphs and runtime relationships from automated tools. As organizations push for higher autonomy and cross-repository coordination, monorepo architecture and synthetic monorepo abstractions are emerging as practical ways to expose complete codebases and persistent knowledge to agents, allowing them to handle more complex, multi-service work with less human orchestration.
Block’s JVM Monorepo: Coordination as an AI-Ready Platform
Block’s migration of about 450 JVM repositories into a unified monorepo shows how consolidating code can reduce friction that stalls AI-assisted work. Under the old polyrepo model, backend services and shared libraries lived apart, leading to dependency version drift, duplicated upgrade efforts, and surprise runtime failures from diamond dependencies. According to Block engineers, the new monorepo now supports roughly 8,800 builds per week, with p90 continuous integration times near 10 minutes on a reliably green main branch. That atomic view is key for AI coding agents: shared libraries are resolved from source, and cross-service changes can ship in a single commit instead of scattered, manually coordinated deploys. Custom tooling, including an IntelliJ plugin that loads only relevant projects and dependency graph–based build scoping, keeps the monorepo responsive so agents and humans can safely run broader refactors without overwhelming CI or the IDE.
Nx Polygraph and the Synthetic Monorepo for Agents
While a physical monorepo changes how teams store code, Nx’s Polygraph aims to give AI coding agents monorepo-like powers without moving any repositories. The service analyzes internal repos and their open-source dependencies to build a dependency graph of which packages publish APIs and which code consumes them. From the agent’s perspective, this graph becomes a single synthetic monorepo: it can read and write across multiple projects as if they lived in the same tree. Co-founder Victor Savkin argues that current agents “run out of stuff to do” because they are bound to one repo and forget everything between sessions. Polygraph counters this with portable shared memory that captures every agent conversation and relates them by intent, not just which files changed. That makes prior work discoverable for new tasks and gives agents context to adjust both producers and consumers of an API in one focused session.
Solving Cross-Repository Coordination and Memory for AI Agents
Both Block’s monorepo architecture and Nx’s synthetic monorepo approach target the same AI limitation: lack of cross-repository visibility and persistent memory for multi-service changes. In a polyrepo world, agents struggle to coordinate updates that span backend services, shared libraries, and client code, because each repository looks like a separate, incomplete system. Monorepos remove those walls by giving agents a single dependency graph and codebase where atomic changes are possible. Synthetic monorepos reach for the same effect across existing repo layouts, giving agents a unified map of APIs and package relationships without forcing teams to migrate. Shared, portable memory means agent sessions no longer reset; instead, traces become organizational knowledge that informs future automation. Together, these strategies push AI coding agents beyond local coding assistance toward reliable cross-repository coordination, where they can address dependency drift, orchestrate refactors, and maintain complex multi-service systems.
Unified Codebase Visibility for Enterprise Agent Orchestration
Enterprise teams experimenting with agent orchestration systems, such as those built on Goose or internal builder frameworks, gain new options when they pair orchestration with monorepo-style visibility. In Block’s case, a cohesive JVM monorepo with clear dependency graphs, merge queues, and selective builds provides the stable substrate needed for agents to propose and execute atomic changes across Cash App and Square services. Similarly, Polygraph’s synthetic monorerepo lets orchestrated agents coordinate work across multiple repositories while drawing on shared memory of past sessions. Instead of agents handing back control after isolated tasks, orchestration frameworks can schedule chains of actions that touch many services but still respect ownership boundaries and CI constraints. As these patterns mature, unified codebase views are likely to become a baseline requirement for serious software development automation, turning AI coding agents into reliable participants in large-scale engineering workflows.
