We instrumented real agent workflows against Throughput Lines of Code Per Day MCP surface: API outline generation, single-function body retrieval, caller mapping, transitive change impact, ranked multi-file search, structural pattern match, context bundle packaging, workspace diff, and attribution traces. The counterfactual is what agents do without a graph — grep plus whole-file reads. On those paths, agents spent roughly an order of magnitude fewer input tokens through outline and query tools than through file dumps. Graph queries are the outliers: callers and diffs return lists of names where the alternative is many whole files.
Blogengineering1 min read
10.7x Context-Token Savings: How We Measured It
Every number on our benchmarks page is copied from measured runs — not marketing fiction.
Token savings claims are easy to invent and hard to defend.
Read the full tables on /benchmarks, then run the same tools against your own workspace.
Continue
Welcome to Throughput Lines of Code Per Day
Agent-native version control that query-maps your repository instead of dumping files into context.
Why Sidecar Code Graphs Go Stale
Sidecars proved agents need a graph. They also proved a second source of truth is a liability.
Plug MCP into Cursor and Claude in 30 Seconds
Zero-friction setup: open Model Context Protocol, one config block, agents start querying structure.