DeepSeek Harness Maintainers

DSH prioritizes UI, composability, and reversibility over agent capability without showing that those investments help the agent.

Yifan Xu’s DSH critique: eight claims, tested separately

DSH inherits much of its core architecture from Cordis and therefore has no genuinely original innovation.

Yifan Xu’s DSH critique: eight claims, tested separately

No public DSH-specific experiment establishes whether DSH prevents unauthorized real side effects under AgentDojo-style indirect prompt injection.

RISK: AgentDojo prompt injection and the untested DSH boundary

DSH tool design mainly splices together existing Claude Code and Codex ideas without its own design understanding.

Yifan Xu’s DSH critique: eight claims, tested separately

Developing with a mature external agent and then reloading into DSH is more convenient and efficient than DSH Creation Mode.

Yifan Xu’s DSH critique: eight claims, tested separately

The DSH Minimalist Benchmark is not meaningful for understanding real harness capability.

Yifan Xu’s DSH critique: eight claims, tested separately

DSH agentic capability is far below the average of comparable harnesses.

Yifan Xu’s DSH critique: eight claims, tested separately

DSH did not publish Code Mode results because its benchmark performance was poor.

Yifan Xu’s DSH critique: eight claims, tested separately

In bounded environments, changing the first-request tool schema materially changes V4 Pro observable reasoning trajectories.

V4 Pro × DSH Minimal: trajectory reproduced, ability unresolved