Research Insights #39 materials: Scion and development with agents (#AI4SDLC)

The materials for the solo episode Research Insights Made Simple No. 39, dated October 15, 2026, are available: “Scion: replacing git inside Anthropic for agent-driven development”. I discussed the company's engineering report "Anthropic — Redesigning source control at Anthropic": what needs to change in code storage when agents and CI generate most of the load.

In the episode, I covered:

🔸 Why an agent needs the whole repository In Scion, the server handles history, merge, and rebase. If an API response is enough for a task, a local copy may not be needed at all. This choice relies on clients remaining connected.

🔸 Why this resembles a database A change log, derived read structures, and caches of immutable objects. Spanner's guarantees, however, cannot automatically be attributed to the entire system.

🔸 What exactly became faster Fetching code and creating a working tree are separate operations. Speeding them up does not yet imply the same improvement across the whole development cycle.

🔸 What the engineers did According to Anthropic's report, seven people built Scion, Antique, and related systems and completed the migration in five months. Humans set architectural constraints and stopped agents from adding unnecessary complexity.

🔸 Where changes are checked now Moving some checks until after merging leaves questions about behavioral compatibility and release approval. The article does not describe the complete path to a verified release.

Anthropic's circumstances are specific: its workload, repository size, and access to models. In the discussion, I separate the implementation described from my own questions about it; this is no ready-made recipe for replacing Git in any team.

Episode materials: 📌 Episode page with timestamps and Episode slides 📖 Longread on infrastructure for agent-driven development 🎬 Video: YouTube and VK Video 🎧 Audio: Podster and Apple Podcasts 📖 Written notes

As the number of agents grows, what hits its limits first in your setup: repository operations, CI, or checking the compatibility of changes?

#AI4SDLC #Agents #Architecture #PlatformEngineering #ResearchInsights