MiniDSH: re-deriving a minimal agent harness from DeepSeek Harness with Claude Code

Beneficial_Owl7036 · reddit · 2026-09-13

To understand agent harnesses at the system level, the author studied DeepSeek Harness (DSH) and re-derived a minimal harness from scratch rather than fork-and-prune, releasing MiniDSH v1.0.

Key architectural properties learned from DSH:

The session/event log as the canonical source of truth
Explicit capability seams instead of one giant Agent object
Authority/policy separated from model reasoning, enforced at the effect boundary
The model treated as a routed capability, not identical to the agent
One runtime under terminal, browser, and headless surfaces
Lifecycle and dependency semantics for components

Simplifications: one package, a tiny composition kernel, sequential foreground execution, bounded one-shot delegation—'minimal surface, complete architecture; fewer concepts, stronger invariants.'

Equally valuable: the Claude Code methodology. To counter AI-assisted development's failure mode of locally-good-but-globally-fragile code, they inverted the process—'architecture defines sessions; sessions do not invent architecture'—deciding architectural home, capability ownership, state boundaries, and dependency direction before adding anything, plus dedicated falsification sessions to overturn stale assumptions. Repo: github.com/earthwalker17/MiniDSH

Original post →

MiniDSH: re-deriving a minimal agent harness from DeepSeek Harness with Claude Code

More from coding & agent