On the same day DeepSeek V4 Pro went to general API availability, the company shipped something arguably more durable: DeepSeek Harness v0.1, an MIT-licensed open-source agent framework that entered developer preview on August 13.
The design thesis is unusually simple to state: everything is a plugin. Not “tools are pluggable” or “you can swap the model backend” — everything. The agent loop itself is a plugin. The UI is a plugin. Storage, security policies, session management, scheduling, context injection — all plugins. The runtime is built on Cordis, a TypeScript meta-framework designed around the idea that software components should be composable both in space and in time — essentially a dependency-injection container where plugins can be hot-swapped at runtime without restarting the process. A running Harness instance is, at the implementation level, just a Cordis context with packages registered against it.
This is meaningfully different from how most coding agent tools are structured. Claude Code, Codex, and similar tools have a fixed core with an extension surface — you can add tools and customize behavior at the edges, but the inner loop is not available for modification. Harness makes the inner loop itself swappable. If you want to replace the orchestration logic with your own, mount a different plugin. If you want to hot-swap a component while a session is running — say, upgrade a model backend without terminating an active session — Cordis handles the reversible effect management that makes that possible without restarts.
The four runtime modes ship out of the box: Standard (full agent with file editing, shell, and web search), PTC (adds TypeScript orchestration for multi-step workflows), Minimal (Bash and string replacement only), and Creation (runtime inspection for building and testing your own presets). These read less like product tiers and more like examples of what configuration-driven composition can produce from the same plugin library.
What makes the release unusual beyond the architecture is the companion research paper. DeepSeek AI and Peking University jointly released A Programming Paradigm for Spatiotemporal Composability, which provides the formal foundations for dynamic component loading and unloading with dependency management. Reversible effects and reactive coeffects are the core mechanisms — they’re what allows Harness to mount and unmount plugins at runtime without leaving the system in an inconsistent state. Publishing academic foundations alongside a developer preview is not how most companies ship agent tooling. Whether the paper lands as useful theory or marketing-flavored formalism is something the broader community will sort out, but the gesture is at least coherent with the “everything is a plugin” claim: if that claim is true all the way down, you need a rigorous account of what makes it safe.
The MIT license is the practical part. The Apache 2.0 and MIT dichotomy has become a proxy battle in the open-source AI space, with some vendors choosing more restrictive terms to preserve commercial leverage. DeepSeek’s choice here means Harness can be redistributed, forked, productized, and built into anything without asking permission or paying royalties. That’s the same choice React and Node.js made, and it’s not an accident that those are the examples cited in the project documentation.
Harness is in developer preview. The Cordis APIs are evolving, the core plugins aren’t stable, and the team is explicit that things will change. But the infrastructure bet it makes is clear: agent frameworks should be composition surfaces, not walled gardens. The next six months will show whether the plugin ecosystem grows around that bet or whether it turns out that most developers want a curated experience more than infinite configurability.