OpenAI rebuilt its Git backend so Codex agents stop waiting on each other

OpenAI says Codex has changed the pace of development inside the company: one engineer now has several changes ready for CI at once, and each push can set several jobs fetching the same commit[1]. From around the end of last year its monorepo systems struggled, with slow fetches and pushes, failed merges, and reliability at times below 90%[1]. The New Stack reports OpenAI expects quarterly commits to the monorepo to be up more than 2,400% by the end of this year compared with the first quarter of 2024[2].

Pando drops two assumptions most Git servers treat as requirements[1]:

• No repository-wide snapshot. Each branch is updated independently, so pushes to unrelated branches don't wait on each other. The tradeoff is that a client reading several branches may see them at slightly different points in time, which OpenAI says users and agents don't depend on.
• No custom packfile per client. Instead of compressing a fresh bundle of objects for every fetch, Pando stores each pushed packfile in blob storage and later hands the same packs to every client that needs them.

Pushes are validated while streaming into blob storage, with a small, constant amount of metadata work per push[1]. Fetches return download links through Git's packfile-uri feature, served by a regional cache service called Express with hit rates above 99%, and OpenAI says clones and fetches on standard cloud VMs can saturate 10 Gbps links[1]. Background compaction merges small packs into larger ones, much like a log-structured database, so a client that is far behind doesn't have to download tens of thousands of tiny files[1].

OpenAI says Pando went from kickoff to production in 10 weeks: one engineer designed the core, teams of Codex agents implemented it in Rust, and it has served production traffic since April[1]. The team wrote a Git conformance test suite so agents could check their changes, used deterministic simulation testing to catch race conditions, and had agents using GPT-6 Astra write Lean and TLA+ specifications to find bugs and prove properties of the system[1].

Pando itself isn't being open-sourced[2]. But OpenAI says several of its client fixes, covering concurrent pack downloads, cleanup after interrupted transfers and resumed downloads, are already upstream, with more in its public Git build, including parallel pack indexing and a new "explicit haves" feature[1]. It is also exploring ways to share the underlying technology with its users[1].

Agentic coding multiplies branches, pushes and fetches, and source control is becoming a bottleneck well beyond OpenAI. The New Stack notes Anthropic has gone the other way with a new client and architecture, while OpenAI bet the problem was how Git is served rather than Git itself[2]. If that holds, teams scaling up coding agents may get relief without retraining anyone off git push.