Experiments, benchmarks, and deep dives from the Xanther team.
We added architectural context to AI coding agents via MCP and tested on SWE-bench Verified. MiniMax M2.5 — a model that costs $0.02 per call — scored 78.2%, surpassing every model on the official mini-SWE-agent leaderboard, including Claude Opus 4.5 at 76.8%.
Django has 4,000+ files and a layered architecture that trips up even experienced contributors. We indexed the entire codebase with XCE and tested against 10 real GitHub issues — agents resolve issues faster, use 57% fewer tokens, and avoid wrong-file rabbit holes.
Coding agents are powerful ships sailing without a map. Context engineering gives them architectural awareness. Without it, even the best models waste tokens. With it, a cheap model outperforms an expensive one.
LobeChat is a 92,000-node TypeScript monorepo. We indexed the entire codebase — 8,764 descriptions, 7,948 LLDs, 796 HLDs — and show how XCE helps agents navigate it in seconds instead of minutes.
Give your agents the context they need to solve real-world issues.