Nx launches Polygraph to help AI agents work across multiple repositories

Nx, the company behind the open-source monorepo build system, has unveiled Polygraph, a platform designed to address a critical bottleneck in AI coding agents: their struggle to manage changes that span multiple repositories.

Polygraph creates what Nx calls a "synthetic monorepo" by stitching together multiple repositories into a unified view that agents can navigate and modify as if working in a single codebase. Co-founders Victor Savkin and Jeff Cross, former Google engineers from the Angular team, argue that current coding agents excel within isolated repos but stall when changes require coordination across dependencies or involve work already completed elsewhere.

The autonomy bottleneck#

According to Nx's modeling, a solo developer using agentic coding tools sees roughly a 4.3x productivity boost, but large organizations—where developers spend significant time coordinating—experience only a 1.3x speedup. The reason: agents run until they exhaust local work, then hand control back to humans. "The limits of agents are set by how autonomous they can be," Savkin notes.

How it works#

Polygraph analyzes a company's repositories and open-source dependencies to build a dependency graph showing which packages are published where and which APIs are consumed by which services. No code moves; Polygraph creates a metadata layer enabling agents to "read everything you want to read and write everything you want to write at the same time, in the same place."

Beyond the unified view, Polygraph includes a shared memory system that captures every conversation an engineer has with an agent. This creates what Savkin describes as "a meta graph of how the work came to be," allowing new sessions to surface relevant past efforts rather than starting blind. Sessions are also portable—another developer can resume someone else's work with full state, repos, logs, and traces intact.

Shipping cross-repo changes#

For multi-repo changes, Polygraph orchestrates the workflow: it sets up affected repos, pushes changes downstream so agents can verify them, and opens pull requests for each. One caveat: PRs still merge separately, not as a single atomic transaction.

Agent-agnostic architecture#

Polygraph functions as a "meta-harness" wrapping around agents rather than replacing them. It currently supports Claude Code, Codex, and OpenCode via the agent-to-agent (A2A) protocol, with Google's Antigravity and others on the roadmap. Because sessions are portable, developers can switch models mid-task if an agent runs out of ideas or tokens.