xAI open-sourcing Grok Build — its terminal coding agent — under Apache-2.0 is less interesting for the “look, another agentic CLI” angle than for what the repo structure quietly admits about running agents in production.

• 🦀 99.6% Rust — the agent runtime, TUI, tool layer, and a workspace abstraction all live in one typed codebase. Memory-safety and a compiler as guardrails for a thing that edits files and runs shell commands on your machine is a defensible choice, not a fashion one.
• 🔌 ACP and MCP as first-class citizens — it speaks Agent Client Protocol for editor embedding and hosts MCP servers, so the harness is a client of your tools rather than a monolith that reimplements each one.
• 🎛️ Three entry points, one loop — interactive TUI, headless for CI, and stdio for editor integration off the same agent runtime. The “one core loop, many front-ends” shape is quietly becoming the default agent architecture.
• 📦 Workspace = filesystem + VCS + execution + checkpoints — treating rollback as a primitive is the part most home-grown agents skip until an edit loop nukes something it shouldn’t have.
• ⚠️ “External contributions are not accepted” — the repo is synced one-way from an internal monorepo. You can read and fork it, but it isn’t a community project. Source-available-under-a-permissive-license, functionally.

The honest read from the HN discussion: this is a reference implementation of a production agent harness, shipped as Apache-2.0, that you’re meant to learn from and vendor — not co-develop. For anyone building an in-house coding agent, a battle-tested checkpoint-and-workspace design you can read beats another blog post about ReAct loops. So is a read-only drop that ships the real architecture more useful to you than a “community” repo that never open-sources the hard parts?