Directed-graph agent cuts tokens 74% with 7回 ➔ 3回 turns

The experiment is a direct test of a design argument by @hanakoxbt: that an agent's inner work should be a small loop of generate-and-verify, while the outer order of stages should be pinned down in ordinary code rather than left to the model. AI Labo built a fully offline sandbox VM in which the AI user and the grading user are separated by Linux UID and systemd namespaces, so the model could not read the hidden test file evaluation/judges.py. On that setup, the single-loop mode spent four of its seven turns on orchestration — inspecting files, running an initial test, re-running tests, and submitting — and its context grew from 1,652 input tokens at the first call to 4,318 by the submit step. The directed-graph mode instead reset context at each node and let a stage-specific verifier advance the workflow automatically, holding input near 2,000 tokens per call. The body frames the takeaway as a separation of reasoning (what to do) from control (how to proceed), and notes that developers could then approve only the final gate once every stage verifier has passed.