Anthropic Replaces Git With Scion as AI Code Strains Engineering Toolchain

Git’s distributed model assumes engineers clone a repository and can commit, merge and inspect history offline. Anthropic’s heaviest readers are CI jobs and agent sandboxes that stay online, run for minutes and are destroyed; they may need a few files, a diff or an ancestry check. Preparing a local repository for each operation became expensive. The main branch history is close to 100 GB and a compressed working copy exceeds 25 GB. Full copies for 25 million daily checkouts would theoretically exceed 600 PB. Shallow clones, sparse checkouts, snapshots and hundreds of Git mirrors reduced the load, but mirrors still needed synchronization, snapshots expired and history queries still required missing objects.

Scion moves commit creation, merging, rebasing and history queries to a central service and leaves a lightweight client that monitors file changes, keeps a small cache and expands a working directory on demand. A CI job that only needs a diff can call an API instead of checking out the repository. Scion records a changeset as a prewrite log entry listing files changed relative to the first parent and references data through content-addressed pointers, so creating a changeset does not require a Git packfile or an immediate move of a shared branch. Blame, path history and commit-graph skip lists are maintained as derived data. During checkout, the client requests only missing trees and blobs through a daemon, a regional cache and a server that falls back to Spanner or object storage. Objects are immutable after write and caches can be reused across tasks; Anthropic says the cache hit rate exceeds 99%. After migration, Scion’s CPU and memory requirements were roughly 200 times lower than the old Git mirror system, checkouts fell from minutes to seconds and CI freed about 25% of CPU capacity. The 200-fold figure refers to infrastructure resource use, not an overall speedup of Git operations. Anthropic also built Antique for code review, source search, IAM and batch merging; running the online service became a new maintenance cost.

GitHub took a different route for secret scanning. Regex finds suspicious high-entropy strings but cannot judge context, and a general LLM can help but adds inference cost and latency as candidates grow. GitHub kept regex as the candidate entry point and moved the decision to a fine-tuned ModernBERT model that reads the candidate string, secret type, variable name, file path and nearby code, then outputs a risk score. The model was trained with hard negatives, strings that look like secrets but are harmless, and with training and evaluation sets split by repository, candidate value and related relationships. It was exported to ONNX and quantized to eight bits. Compared with the old LLM pipeline, offline precision rose 13% and recall rose 19%. In online A/B tests, median post-push scan execution time fell 86%, model serving cost fell more than 98% and a repository-owner-weighted alert acceptance metric rose 11.5%. The 98% figure covers model serving, not total security-system cost. The classifier only judges candidates found by regex, so rule coverage still limits missed detections.

Researchers at Columbia University, Google DeepMind and other institutions built BTTF to make chip simulation waveforms queryable by agents. RTL simulation produces VCD and FSDB files that record signal states over time; complex dumps can reach tens of gigabytes or terabytes. Converting raw waveforms to text fills an agent’s context with timestamps, signal names and repeated transitions. BTTF loads the data into SQLite. A signal_metadata table stores names, types and hierarchy, while signal_changes records signal ID, timestamp and new value, with the two linked by ID. An agent can ask SQL for the most recent change before a given time or filter and aggregate transitions in a time window, then inspect only a few returned events. BTTF maps waveform identifiers to database IDs during ingestion, counts transitions and applies clock filtering, no-toggle signal pruning and batch inserts.

An untrimmed database took 1,420 MB and queries took about 410.5 milliseconds; filtering clocks and non-toggling signals cut the database to 395 MB and query time to 94.8 milliseconds. Deleting signals with below-average toggle rates reduced the database to 185 MB but left activity coverage at 64.23%, and rare faults can occur on control signals that almost never toggle, so pruning has to follow assertion and debugging needs. With a SQL interface, BTTF assigns roles to different agents: an orchestration agent decomposes tasks, an analysis agent writes queries, a verification agent checks assertions and links RTL source, and an evaluation agent checks time ranges, field relationships and result completeness. In an experiment with a 4.2 GB waveform, 18,450 signals and 150 expert-annotated queries, BTTF reached 95.33% execution accuracy. Multi-agent quality scores were 17.3% higher than a single agent, average latency rose to about 1.96 times, and static SVA checking accuracy was 84.62%. The system can support post-simulation analysis, but continuous ingestion and complex attribution still need validation.

Anthropic’s source-control fix exposed the next bottleneck. A test run may still be in progress when the main branch merges hundreds or thousands of additional changes. A global lock lets changes pile up; loosening concurrency creates risk that a passing test no longer matches the code state being prepared for deployment. CI therefore has to keep validation results valid under frequent change: build caches need to detect source and dependency changes, test records need to be tied to specific commits and build artifacts, and the system must decide which results can be reused and which changes invalidate earlier verification. In complex monorepos, adding test machines does little for version consistency under concurrent edits. Anthropic has strengthened post-merge testing and pre-deployment validation. For ordinary teams, improving dependency analysis, incremental testing and build caching is more realistic than rebuilding the version control system first, according to Leiphone.