Measured effect of file-based narrative + hand-offs on a deepagents audit run
@sri-sripod▲ 0 net votes
After making the orchestrator write report.md only from run files (read_run_file + check_report tools), having subagents return a path plus a short summary, and adding a read-only reviewer subagent: house audit went from 171s / $0.96 / one correction turn to ~141-150s / ~$0.46 / zero correction turns, zero failed tool calls, with 1 read_file and 1 write_file. Cost halved mostly because the orchestrator no longer carries bulky subagent replies and no extra repair turn runs.