DataFlow-Harness cuts pipeline-agent cost 72.5%

DataFlow-Harness reports a 93.3% end-to-end pass rate across 12 data-engineering tasks while cutting measured construction cost 72.5% versus vanilla Claude Code. The technical report describes an agent constrained to typed DAG mutations, live operator metadata, and procedural skills rather than disposable scripts; the Apache-licensed implementation cautions that its benchmark inputs and raw results are not included for independent reproduction. For builders weighing MCP’s context tax against governable agent workflows, the result is worth filing as a design signal—not a verified universal speedup: constrain the artifact and expose only the platform state the agent needs.