no-graphvscode graph  · agent benchmark

The same unrestricted agent, run on the same graded questions with and without a code graph. Arm A has no code graph; Arm B has the auryn MCP server. Measured on tokens, cost, wall-time, and key-fact coverage. Click any arm to replay the agent's full trace.

A / B means, all tasks
Per task — click an arm to open its trace
drop run JSON files here (a results/<stamp>/runs/*.json set) to load another benchmark