DeepCode v2 — Loop Engineering for Agentic Coding
LinkedIn post by Chao Huang (leads the Data Intelligence Lab @ HKU; nanobot, LightRAG, RAG-Anything, DeepTutor, AI-Trader…) — view post
Problem statement: “An agent that can write code doesn’t always finish the job” — complex tasks stop halfway, verification gets skipped, and regression checks + repo maintenance still need repeated manual prompting.
What v2 ships
- 🔄 Loop Engineering — give it a natural-language Goal and it keeps working through repository understanding → implementation → testing → verification → repair, instead of stopping after one response.
- 🎛️ Human stays in control — add requirements, revise the Goal, pause/resume/redirect at any time; completed work and context are preserved.
- ✅ Verified delivery — runs commands, tests, builds, runtime checks; connects every code change with execution evidence so results are easier to review and merge.
- 🤖 Personalized automations — project-specific natural-language tasks run manually or on schedule (regression testing, test repair, doc updates, repo maintenance).
- 💻 One experience, CLI + Desktop — same Project/Session/Goal/execution history from either interface.
Positioning: not just a code generator — an open-source general-purpose coding agent designed to keep working inside real repositories until complex tasks complete.
- Repo: https://lnkd.in/gJvh6qXi (GitHub
HKUDS/DeepCode— “Open Agentic Coding: Agent Harness & Loop Engineering & Multi-Agent Orchestration”)