DeepSeek Harness Explained: The Fully Modular 'Everything is a Plugin' Agent System
数字生命卡兹克 · wechat · 2026-08-14
An in-depth look at DeepSeek's newly released Agent product, DeepSeek Harness. Its core philosophy is "everything is a plugin," designed to allow Agents to self-evolve by dynamically managing plugins during runtime.
- Core Architecture: Built on the Cordis kernel (developed by a newly acquired DeepSeek team member). It is highly restrained, handling only plugin loading, unloading, and dependency management. It features Temporal Composability (completely reverting side effects when a plugin is unloaded) and Spatial Composability (dynamically readjusting dependencies).
- Four Working Modes:
- Standard Mode: Pre-configured with full coding agent capabilities, ideal for beginners.
- PTC Mode: Provides a Code Mode SDK, allowing the model to write TypeScript programs to execute multiple tool operations in a single run, reducing model round-trips and saving tokens.
- Minimal Mode: Strips down to a persistent Bash and file editor, specifically for benchmarking a model's raw agent capabilities.
- Creation Mode: The standout feature, allowing the Agent to inspect its own environment and create/test/install new plugins on the fly, effectively modifying itself.
- Developer Experience: Conversations are designed as an append-only event log, making agent trajectories fully observable, auditable, and reproducible for research.
- Pricing & Models: Supports integrating third-party models. DeepSeek V4 Pro's official pricing has significantly increased (cache hit price up 12x, peak output at 27 RMB/million tokens), reducing its cost-effectiveness edge.
- Ecosystem: A community plugin marketplace launched alongside the product. The author notes the steep learning curve for average consumers but highlights its massive value for research and exploration.
Related event: DeepSeek Open Sources Plugin-Based Agent Framework Harness v0.1(36 posts)→
More from coding & agent
- edith-1 monitors agent traces, beats Sonnet-5.5 on balanced accuracy at 1/274 the cost — xennygrimmato_ · 2026-10-09
- How edith-1 works: probabilistic filter plus expensive agent judge for flagged runs — xennygrimmato_ · 2026-10-09
- Splash 1.3.0 cuts local agent first-token latency from 19s to 1s via SSD offloading on M6 Mac — BeidiChen · 2026-10-09
- OpenAI makes coding-agent steering instant, ships new "GPT-6.1 Sol ultrafast" release — Dimillian · 2026-10-09
- Robotics dev: LLM users keep spawning new robot libraries instead of improving existing ones — kscottz · 2026-10-09
- Modal Launches Sidecars: Sandboxed Containers for Agent-Generated Code with 3x Faster Cross-Trust-Boundary Communication — AAAzzam · 2026-10-09