DeepSeek Harness developer preview
TL;DR Highlight
DeepSeek released a plugin-based agent harness as an MIT-licensed developer preview.
Who Should Read
Developers using coding-agent harnesses like Claude Code or Pi who want a cheaper model option.
Core Mechanics
- The source is only a redirect notice
- The article text says just 'Page moved Continue', with no feature details.
- Product details come only from comments, so we don't state them as fact here.
- The author calls it an early preview
- A commenter who says they are an author says it is an early MIT-licensed developer preview.
- They warn of rough edges and breaking changes.
Community Reactions
- Full session logs praised
- A quoted description says system prompts, reasoning, tool calls and subagent scheduling go into an append-only log.
- One commenter said US models encrypt their traces, making this kind of analysis hard.
- Cordis plugin architecture explained
- Commenters say it runs on Cordis v4, which loads and unloads plugins without a restart.
- On unload it reportedly reverts state and side effects and cleans up dependencies.
- Plugin fatigue and sparse README
- One commenter fears 'everything is a plugin' ends in incompatible, deprecated plugins.
- Another said the README is bare and doesn't explain what it is.
- Local model run reported working well
- One user ran it with Bun and a local 9B Qwen model on llama.cpp for small Python projects.
- They found it quick compared with other harnesses.
How to Apply
- Install with Bun and connect a local model
- Follow the install steps in the GitHub repo today.
- Give it one small Python task and compare against your current harness.
- Inspect the Trajectory view
- Run the same task, then review tool calls and context injections by source.
- Note whether failures come from the prompt or the tools.
Terminology
Related Papers
Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy
Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.
Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.
Harness Engineering 101: The LLM is the brain, the harness is the body
An agent is a loop managing a JSON array; 16 chapters build one from scratch
Harness2 by Google Research – Recursive Agent Harnessing
Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.
Ousted Savannah police officer gave ICE agent husband access to Flock system
A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
Running Claude Code agents in tmux works well, but limits appear around four agents.