MiniDSH: re-deriving a minimal agent harness from DeepSeek Harness's architecture
Beneficial_Owl7036 · reddit · 2026-09-13
The author studied DeepSeek Harness (DSH) and re-derived a minimal agent harness from scratch — MiniDSH, now at v1.0 and MIT-licensed — to answer what architectural properties a real-world harness minimally needs. Key ideas carried over: session/event log as source of truth; explicit capability seams instead of a monolithic Agent object; authority/policy separated from model reasoning and enforced at effect boundaries; the model as a routed capability; one runtime under terminal/browser/headless surfaces; lifecycle and dependency semantics for components. He aggressively simplified elsewhere (single package, tiny composition kernel, sequential foreground execution, bounded one-shot delegation). He also describes building it with Claude Code, inverting the usual failure mode of AI-assisted development: architecture defines sessions, not the other way around, with dedicated sessions to falsify architectural assumptions — several original plans were overturned by evidence.
More from coding & agent
- ProTip: lock down code sections in agents.md to stop agents from breaking them — cyrus_zei · 2026-09-13
- Open-source thesys-core highlights exact paragraphs behind AI answers in 100+ page PDFs — Flat-Phone-1596 · 2026-09-13
- Dev uses idle Opus 5 credits to build split-screen multiplayer into his game — gandamu_ml · 2026-09-13
- freecad-mcp: 2.2k-star MCP server lets Claude Desktop drive FreeCAD for CAD and FEM — tom_doerr · 2026-09-13
- Dev fine-tunes Qwen 3.8 on 81,837 book annotations, writing quality up 86% in blind tests — Scobleizer · 2026-09-13
- Connecting health wearables via API/MCP gets tangled: use Whoop as single source of truth — gethackteam · 2026-09-13