Artificial Intelligence

DeepSeek Harness Is a Developer Powerhouse, but Casual Users Should Wait

By 深小乐 · Aug 19, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The plugin-everything model and self-authoring agents point toward a future where AI tools extend themselves without human intervention, but the current build is too raw for anyone who isn't comfortable debugging agent behavior and restarting services manually.

Summary

DeepSeek Harness ships with a plugin-everything architecture that lets agents discover missing functionality, locate a plugin, install it, and restart the service without leaving the conversation. The agent can also write new plugins from scratch, as demonstrated by a clipboard-history manager built entirely through prompting. A detailed execution trace shows every tool call, token burn, and failure point on a color-coded timeline, giving developers full observability into agent behavior.

The community has exploded: over 6,600 plugins and 147K GitHub stars within four days of release. Token costs are remarkably low, with cache-hit rates reaching 98%, though DeepSeek has announced a price increase is coming.

Version 0.1 is explicitly labeled for Harness developers. The web UI is bare-bones, lacks multimodal input, has no scheduled-task support, and the plugin directory shows only names with no descriptions. UI bugs during plugin development required over an hour of back-and-forth debugging, and the model's inability to accept screenshots made troubleshooting harder than it needed to be.

Takeaways

— DeepSeek Harness treats every capability as a plugin, including the ability for agents to find, install, and write plugins themselves.

— Execution traces expose every tool call, context window, token cost, and failure on a color-coded timeline, similar to a network inspector.

— An agent successfully located and installed a Mermaid rendering plugin after the user asked for a diagram it couldn't display natively.

— Building a clipboard-history manager plugin through prompting took over an hour due to repeated UI bugs and the model's inability to accept screenshots.

— Token costs are extremely low with a 98% cache-hit rate, but DeepSeek has announced prices will rise.

— The community produced 6,600+ plugins and 147K GitHub stars within four days of launch.

— Version 0.1 is explicitly labeled for developers; the UI lacks multimodal input, scheduled tasks, and meaningful plugin descriptions.

Conclusions

Self-authoring agents that write their own plugins blur the line between tool-user and tool-builder, but the debugging overhead shown here suggests the workflow still demands a developer's patience and troubleshooting skill.

The absence of multimodal input is a practical bottleneck: describing a UI bug in text took multiple rounds, where one screenshot would have sufficed. Competitors like Codex already support this.

A 98% cache-hit rate on agent sessions implies the architecture reuses context aggressively, which keeps costs down but may also explain why the agent struggled to escape its own flawed fixes during debugging loops.

The explosion of community plugins in under a week signals intense developer enthusiasm, but a directory with name-only listings will quickly become unusable without curation or search improvements.

Concepts & terms

Execution Trace (Trajectory)

A detailed, time-ordered log of every action an AI agent takes during a session, including tool calls, token consumption, context state, and failure points, displayed with color-coded blocks for inspection.

Plugin-Everything Architecture

A design where every capability of an AI agent platform is implemented as a plugin, allowing agents themselves to discover, install, configure, and even author new plugins at runtime.

Cache-Hit Rate

The percentage of LLM inference requests that can reuse previously computed results instead of running a full forward pass, directly reducing token costs and latency.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗