What I Learned Running DeepSeek Harness: an Open, Plugin-First Agent Runtime

# Is DeepSeek Harness (CLI: dsh) is an open agent runtime released by DeepSeek in developer preview. Its central idea is simple and strict: every component of an agent is a plugin. That includes model adapters, tool registries, session logs, sandboxes, UI panels, and even the agent loop itself. The project is built on Cordis, a plugin framework with prior production use inside the Koishi chatbot project.

# Why people noticed it quickly The repository received extremely fast attention after release: roughly 50,000 stars in the first 12 hours, about 92,000 by hour 28, and over 186,000 within ten days. That velocity reflects two things visible in the code and docs: a real, fine-grained plugin architecture and transparent, documented design (including an accompanying paper, A Programming Paradigm for Spatiotemporal Composability).

# Concrete technical choices that matter

• Model-agnostic: dsh supports about 40 model providers and allows delegation of sub-agents to external agents rather than locking users to DeepSeek models.
• OS-level sandboxing: uses bwrap and Landlock on Linux, Seatbelt on macOS, and restricted ACL tokens on Windows. Sandboxes are fail-closed by default.
• Append-only session logs: runtime-enforced logging policy ensures anything the model saw is recorded.
• Plugin surface is fine-grained: UI pieces, approval prompts, sub-agent panels, and scheduler components are separate installable packages.

# What I ran and what I saw

# What this product is right now DeepSeek labels the repo "developer preview" and warns explicitly about compatibility-breaking changes. This means:

• It's infrastructure for building agents rather than a finished agent product.
• Expect breaking changes and active development.
• The architecture is intended for teams that want modular, auditable agent components and strict sandboxing.

# Why that matters for engineers If you need control over agent composition — swapping model providers, isolating tool execution at the OS level, or auditing session context — Harness makes those parts explicit and replaceable. Its architecture can reduce lock-in: sub-agents can call out to other agents, and plugins are regular packages you can fork or replace.

# Practical next steps if you want to try it

• Install via npx to confirm the CLI and version.
• Use --dump-default-config to inspect a profile's plugin tree.
• Start with the web or headless profile to see how UI and headless task execution differ.
• Treat it as a platform to assemble and test agent components rather than a ready-to-run coding assistant.

# Bottom line DeepSeek Harness is a transparent, plugin-first agent runtime designed for teams that want modularity, strict sandboxing, and auditable logs. It's not a completed, polished agent product yet, but it provides the building blocks developers can use to assemble agents under predictable, replaceable pieces.