DeepSeek harness: open-source agent with an inspectable trajectory view

Melvin Vivas · X post · 2026-08-19 · Open on X

Topics: AI Agents, Tool Use & MCP, AI Dev Tools & Productivity, Evaluation (Evals) & Testing · Level: advanced

Summary

The creator highlights the DeepSeek harness, an MIT-licensed open-source agent harness. Everything in it is a plugin, it works with any model, and it has a detailed trajectory view of the event stream. He suggests collecting these traces so a harness can improve itself based on how you use it.

Key points

• DeepSeek harness is fully open source under the MIT license.
• Everything is a plugin that you can customize on the fly by prompting.
• The trajectory view shows a very detailed event stream, so you can inspect what the agent did.
• It works with any model (GPT, Claude, etc.).
• Idea: collect the traces and use them to make your harness self-improve based on your usage.

Resources mentioned

• DeepSeek Harness · repo · github.com · free
  Open-source (MIT) agent harness with plugins, a trajectory/event-stream view and support for any model.
  Also in: DeepSeek Harness desktop app released for macOS and Windows (Melvin Vivas on X · notes), One-Shot Agent Management UI with DeepSeek Harness for $0.19 (Melvin Vivas on X · notes), Getting Started with DeepSeek Harness and the DeepSeek API Platform (Melvin Vivas on X · notes), DeepSeek Harness: Open-Source, Browser-Based Agent for DeepSeek V4.1 (Melvin Vivas on X · notes)

Try this

• Try the DeepSeek harness and look at its trajectory view.
• Collect agent harness traces and build a loop that improves the harness from your own usage.