Coding agents are forcing a rebuild of the whole dev stack — from version control to evals

The plumbing

The week's loudest infrastructure story: Anthropic replaced Git on its giant monorepo with a brand-new server-side source control system called Scion. "Anthropic's monorepo grew from 2,500 merges a day in May to over 30,000 in September" — and "Scion system uses 200x less CPU and memory than the fleet of Git mirrors it replaced." Alongside it came Antique, a review system where engineers review the intent of changes rather than reading every line. Anthropic does not intend to open-source Scion, but it is offering to work with the open-source community on server-side APIs for Git history and diffs. GitHub, meanwhile, is rebuilding its own Git infrastructure by separating storage from compute on the server — the same architectural instinct, applied at cloud scale. m:…6931

At the IDE layer, GitHub announced that Copilot app sessions can be opened directly in VS Code, preserving the task, plan, and conversation history. You keep prompting in the sidebar, manual edits build directly on the agent's work, and changes push to the associated pull request — the handoff between the agent's workspace and the human's editor, which used to mean copying files between directories, is gone, the company says. m:…9145

The scale that forced the rewrite is worth sitting with: Anthropic ran roughly 25 million CI checkouts a day, its main-branch history was nearly 100GB, and it moved to testing after merge and before deploy to keep up with concurrent changes, according to the same report. At that volume, version control stops being plumbing and starts being a first-class infrastructure bet. m:…6931

The training data

Agents trained on toy repos break on enterprise ones, so the corpora are growing up. In a preprint, the CIDR dataset collects 1,033 million raw lines of code across 4,687 industrial repositories, tagged with engineering-practice attributes like CI usage and automated test presence. Fine-tuning a 3B-parameter model on it improves performance on held-out enterprise code — a small-model result with a practical reading: your training distribution is the product. m:…4697

The evals

The eval layer is being rebuilt around dependency structure instead of flat scoring. In a preprint, WEAVE transforms verified PR histories into executable, dependency-structured tasks, and its Graph Credit metric gives a deterministic, test-grounded way to measure partial progress on multi-PR work — the authors argue it measures long-horizon progress more faithfully than pass/fail. m:…8309

EvoWeb attacks the problem from the other end: in a preprint, it benchmarks agents on 320 cumulative requests over 32 evolving web applications, and finds that cross-request integration is the key challenge for frontier models — later requests with deeper or more numerous dependencies are solved less accurately. Your agent's fifth change to a codebase is harder than its first, and now there's a number for it. EvoWeb also integrates trajectory performance across user-tolerance thresholds instead of using a fixed stopping criterion, the authors say — and runtime errors become more frequent over successive requests, which is exactly the degradation pattern long-horizon agent teams need to budget for. m:…4916

And SWE-Prometheus, in a preprint, measures what most evals ignore: repo governance. Repository-blind templates improve tests, CI, quality gates, and documentation — but fail to improve reproducible environments or dependency and security hygiene, with behavior-breakage rates for agents ranging from 0% to 23%. If your team ships agents into repos, the eval question isn't just whether the code works; it's whether the repo still does. m:…0273