DeepSeek
Open-source LLM optimized for advanced reasoning and code
4.9•49 reviews•3.5K followers
Open-source LLM optimized for advanced reasoning and code
4.9•49 reviews•3.5K followers
Intelligent assistant for coding, content creation, file reading, and more. Upload documents, engage in extended conversations, and receive expert assistance in AI, natural language processing, and beyond.
This is the 15th launch from DeepSeek. View more
DeepSeek Harness
Launching today
DeepSeek Harness is an open-source agent runtime where models, tools, prompts, storage, the agent loop, and even the UI are plugins. You compose profiles and agent presets, run programmatic tool calling, and keep a full event log for recovery and replay.
Free
Launch Team
ChatGPT by OpenAI
Flowtica Scribe
Hi everyone!
It’s easy to look at this and think it’s just "@DeepSeek’s version of Codex." But DSH is actually a recomposable agent runtime rather than a fixed product.
The core idea is everything is a plugin. The model, tools, approval policies, context handling, and even the web UI itself are swappable components you can compose or replace. During the closed beta, devs built fully custom interfaces and long-term memory systems entirely through plugins, without forking the core code.
Agent Presets let you assemble very different agents from the same runtime. There’s also PTC, where the model can write a TypeScript program that combines multiple tool calls into one execution, keeping intermediate data out of the model context and cutting down a lot of back-and-forth.
It defaults to DeepSeek V4 Flash, but you can plug other models into it too.
DSH exposes the layer between the model and the finished agent, and makes almost all of it configurable.
@zaczuo the line that got me was PTC — letting the model write a TypeScript program that fans out several tool calls in one execution, so intermediate data never touches the context window.
I spend a lot of time on agent reliability, and most of my context bloat isn't reasoning, it's tool payloads I never needed the model to read. Solving that at the runtime layer instead of with yet another summarizer is the right place to solve it.
The full event log for recovery and replay is the other half people will underrate until their first bad prod run 👌