Author(s): Udaykiran Estari

Originally published on Towards AI.

Stop comparing marketing pages. We swapped the models underneath both agent harnesses to isolate the architecture from the hype.

Every launch-week take treats the new DeepSeek Harness versus Claude Code as a proxy war between their underlying language models. This conflation masks the most important architectural shift in agent development: harness quality and model capability are two separate variables, and conflating them leads to expensive, rigid architectural decisions. If you want to know whether DeepSeek’s “everything is a plugin” approach actually threatens Anthropic’s integrated developer experience, you have to stop comparing marketing pages and start swapping components. We dissected Cordis’s effect/coeffect plugin contract line-by-line and literally swapped brains — running Claude inside DeepSeek Harness and DeepSeek V4 inside Claude Code — to show you how to isolate the harness from the model.

The article argues that “harness” and “model” are different axes, so meaningful comparisons require isolating each. It explains what a harness is (scaffolding for tool use and orchestration) and why DeepSeek’s “everything is a plugin” claim is grounded in Cordis’s effect/coeffect mechanism: plugins declare dependencies and only activate when satisfied, while their registered capabilities are treated as reversible effects with no privileged core. To test the separation in practice, it reports two controlled swaps: running Claude’s model inside DeepSeek Harness by changing only the model adapter configuration, and running DeepSeek’s V4 under Claude Code by using an Anthropic-Messages-compatible endpoint with server-side model remapping. The results suggest the “feel” of a harness remains consistent across models (the harness personality is sticky), while the model’s judgment shifts outcomes. The conclusion weighs cost and capability—DeepSeek’s pricing can make heavy parallel agent workloads dramatically cheaper, though top-tier single-chain reasoning may still favor Claude—while warning that Harness is still developer-preview and inherits prompt-injection risks when executing shell commands, so isolation (low-privilege VM/container) is essential.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI