DeepSeek just put its agent harness into developer preview, MIT-licensed and positioned squarely against Claude Code. The headline is the open source. The more interesting signal is what the release asserts: at this point the harness — not the weights — is where agent capability is won or lost.
- 🧩 Everything is a plugin — models, tools, skills, sessions, sandboxes, loops, and the UI all mount on a “Cordis” kernel, so you can swap any layer without forking the whole stack.
- 🎯 The harness is the product — DeepSeek is drawing the line where I draw it in production: the model is one replaceable dependency; the orchestration around it is the durable asset.
- ⚠️ Developer preview means churn — open code, but breaking changes are promised. This is something to study and prototype against, not to standardize a team on yet.
- 🔍 Plugin seams cut both ways — clean boundaries for composition also mean a bigger untrusted-code surface (remote plugins, sandboxes) whose security you now own.
- ⚡ Portability is the actual pitch — if tools and skills are plugins, they stop being locked to one vendor’s agent loop, which is exactly what enterprise buyers keep asking for.
Early coverage splits along the preview-versus-production line. Crypto Briefing reads the MIT license as a shrewd, developer-friendly shot at Claude Code; explainx.ai agrees the swappable-component architecture is technically substantive but says the breaking-change warnings make it a learning tool, not a production platform yet; VentureBeat frames it as an open-source rival while noting the paired V4-Pro API landed at higher prices. The HN thread tracks the same tension — real enthusiasm for the plugin model, real wariness about betting live workflows on a v0.1. That’s the right read: the idea is worth stealing even if the release isn’t yet worth adopting.