DataFlow-Harness uses MCP and DAG edits to build editable LLM pipelines
PekingUniversity · hf · 2026-07-22
Peking University’s DataFlow-Harness aims to close the NL2Pipeline gap: coding agents often generate scripts, but not persistent, editable pipeline artifacts.
Approach
- Guides an LLM agent to build platform-native DAGs through typed, incremental mutations instead of free-form scripts.
- Adds DataFlow-Skills for procedural guidance.
- Uses an MCP layer to expose the live operator registry and current pipeline state.
- Includes DataFlow-WebUI, which keeps conversational authoring synchronized with a visual DAG editor.
Results
- On a 12-task data-engineering benchmark, it reports a 93.3% end-to-end pass rate.
- Compared with Vanilla Claude Code, it cuts measured cost by 72.5% and latency by 49.9%.
- Its pass rate is within 0.9 points of the Context-Aware Claude Code baseline, while still costing 42.8% less.
More from coding & agent
- ISO says RLVR can keep the weight spectrum and learn 2.7× faster — UTEXAS · 2026-07-22
- AI agents raise a new payment question: cards, bank transfers, or stablecoins? — sweetheartcheek · 2026-07-22
- Codex with img2threejs beat Blender MCP in a quick 3D workflow test — Dimillian · 2026-07-22
- Claude Code task planning plus subagents cuts token costs, user says — toptickcrypto · 2026-07-22
- Image MCP Server brings OpenAI and Replicate image generation into agent workflows — modelcontextprotocol · 2026-07-22
- agentdata-mcp exposes crypto prices, yields, and arbitrage tools to AI agents — modelcontextprotocol · 2026-07-22