DeepSeek Harness Just Changed AI Forever

EnglishالعربيةDeutschEspañolFrançaisहिन्दीBahasa Indonesia日本語한국어PortuguêsРусский中文
BBetter Stack
Computing/SoftwareInternet Technology

Transcript

00:00:00deep seek just dropped their own custom harness and it's skyrocketed to 150 000 stars on github
00:00:05that is a lot of open claw instances it's open source and mit license so if you want to fork
00:00:10the repo just to tell everyone you work in ai then you can do that right now and the plugin
00:00:15system for this is insane letting you customize literally anything you can imagine so let's
00:00:20download the old boy and see how it runs so docker sorry i mean deep seek no wait the one with the
00:00:30blue whale mascot are killing it right now they've just dropped v4 pro which is sitting with the top
00:00:35models at a fraction of the cost you can see it matches glm 5.2 on intelligence but comes in
00:00:41slightly cheaper and it's not far off opus considering the absolute insane price difference
00:00:46and now they're dropping their own custom harness it's built on top of cordis kernel and if you've
00:00:51never heard of that before then no worries it's very well explained on the github page
00:00:55a meta framework of spatio-temporal compatibility obviously it just rolls off the tongue and cordis
00:01:01isn't something deep seek not together for this it's been sat on github since 2022 so they've taken
00:01:07a plugin framework that already existed and built a whole harness on top of it because literally every
00:01:12agent capability inside deep seek harness is a plugin meaning it's highly customizable and while spatio
00:01:19temporally composable but it's actually pretty cool literally everything from the models the tools the
00:01:24skills and everything else listed here are all provided via plugins from cordis which means you can just
00:01:30turn the most core pieces of functionality off and completely lobotomize the application and because
00:01:35the model itself is just a plugin you're not even tied to deep seek's own models their docs walk you
00:01:40through adding anthropic open ai bedrock or any open ai compatible endpoint you've got lying around their
00:01:47own config example even has claude sonic 4.5 sat in it they even say developers can select swap or extend
00:01:53any capability in configuration without changing the deep seek harness source code and we are going to
00:01:58touch on plugins much more deeply in just a second but this next feature i really love and it's traceability
00:02:04built in this is an append only log that records everything the model sees including system prompts
00:02:09reasoning and tool calls so you can see a full history of absolutely everything we're here inside
00:02:15the deep seek harness and in this case i'm using the desktop app at the top here we've got chat and
00:02:20trajectory if i click into this i can see a full history of everything that's happened in this entire
00:02:25conversation i can also highlight areas as well if i want to zoom in on a particular section we can also
00:02:30search up here for any words that were mentioned this also covers the entire conversation the tool calls
00:02:37the thinking steps so let's find let's click on this one for example you can see this took 187 tokens
00:02:46the reasoning was 94 of those tokens and i can also click into the thinking steps to see all of the
00:02:51models thinking and we can do this for every single part of the conversation so that is pretty powerful
00:02:57now let's jump back over to the plugin system because this is the biggest feature of the entire harness in
00:03:01my opinion and if you do want to develop your own plugins you can easily do that within the harness
00:03:06itself we're given multiple runtime modes for this standard mode includes the full tool set and you'd
00:03:12want to use this on your day-to-day engineering tasks but creator mode gives you extra capabilities to
00:03:17inspect and author plugins so you don't have to leave the deep seek harness at all to customize it so
00:03:23let's take a look at customizing this ourselves so to add a new plugin inside chat i've sent an initial
00:03:28message add a new plugin to add a dinosaur jumping game in the bottom right of the application i'm in
00:03:34creator mode here so it has access to all of the tools to be able to execute this and the result is
00:03:39that we get this cool little dinosaur in the bottom right i can click it and now we can start playing
00:03:44the game nice way wait wait yes boy so basically any functionality that you want to add to the
00:03:54application can all be done through a prompt so deep seek can add plugins to its own harness they also
00:04:01provide a bunch of examples on their marketing site like adding a floating whale or a snake game and this
00:04:07is genuinely one of the coolest features i've ever seen inside a harness open code also provides similar
00:04:12customization through prompting but the user experience here and the level of customization
00:04:17all through plugins is really really smooth and because the whole thing is open source people have
00:04:22really been running with this within a few days of the repo going up there's a desktop app which is
00:04:27actually the thing i'm using right now a clawed code style tui and a pack of web ui skins and an awesome
00:04:33list none of which are actually built by deep seek themselves the desktop app is on 11 000 stars by itself
00:04:40if we look inside settings as well we you see i have plugins here we've got a massive plugins list so
00:04:46literally everything the llm the sessions the api gateway the agents themselves are all provided through
00:04:53plugins so you can change this application in any way you could possibly imagine all through a single
00:04:59interface but other than that you get all of the standard functionality that you would expect from a harness but
00:05:04again all provided through plugins you can send chat messages called tools interact with code on your own
00:05:10machine so anything that you could basically do include code you can also do in the deep seek harness but
00:05:16look this is so extensible i can just see it being an extremely powerful tool it's in beta right now so
00:05:21there are a few problems with it but that's expected but i can really see this growing into something
00:05:26special within the next 12 months i've left a link to the repo down in the description you can check it
00:05:31out there and if you want to check out another harness i've covered qm from y combinator a couple
00:05:35of weeks back otherwise go and check out deep seek and maybe try and author some plugins let me know
00:05:39what you think in the comments thank you so much for watching i've been warren from better stack and
00:05:44i'll see you in the next one

Key Takeaway

DeepSeek Harness provides a fully open-source, highly extensible agent architecture built on the Cordis kernel, enabling developers to customize models, tools, and UI interfaces entirely through plugins and natural language prompts.

Highlights

  • DeepSeek Harness reached 150,000 GitHub stars under an open-source MIT license.

  • The architecture uses the Cordis kernel framework, built upon a meta-framework of spatio-temporal compatibility from 2022.

  • Every agent capability inside the harness functions as a plugin, allowing users to turn off core functionality or swap models like Claude Sonic 4.5.

  • Traceability logs record append-only histories of system prompts, reasoning tokens, and tool calls with fine-grained inspection steps.

  • Creator mode enables users to author and add functional UI plugins, such as a dinosaur jumping game, purely through natural language prompts.

Timeline

DeepSeek Harness Introduction and Architecture

  • DeepSeek released a custom open-source harness under the MIT license.
  • The repository achieved 150,000 GitHub stars shortly after launch.
  • The system is built on top of the Cordis kernel framework originating from 2022.

The release introduces a modular harness designed for agent capabilities. It leverages an existing plugin framework known as Cordis, utilizing spatio-temporal compatibility principles to manage core modules. Users can fork the repository or integrate alternative models such as Anthropic, OpenAI, or Bedrock endpoints alongside native DeepSeek options.

Plugin System and Traceability Features

  • Developers can swap or extend any capability without altering source code.
  • An append-only log records system prompts, reasoning steps, and tool calls.
  • Individual conversation tokens and thinking steps are inspectable directly within the desktop interface.

Every core piece of functionality operates as a plugin, granting developers total control to modify or lobotomize application components. The traceability feature records a transparent history of interactions. Users can inspect specific token expenditures, such as reasoning token breakdowns during tool executions.

Creator Mode and Custom Application Extensions

  • Creator mode gives users runtime capabilities to inspect and author plugins directly within the app.
  • Natural language prompts allow the application to generate functional mini-games like a dinosaur jumper.
  • Community developers built independent desktop apps, TUI interfaces, and web UI skins.

The platform includes specialized runtime modes for engineering tasks and plugin authoring. Prompting the chat interface enables the model to write and install its own plugins, such as interactive UI elements or custom games. The open-source nature prompted the community to rapidly build alternative wrappers and toolkits.

Ecosystem Extensibility and Future Outlook

  • LLMs, sessions, API gateways, and agents are all provided through modular plugins.
  • Standard harness features include messaging, tool calling, and local code interaction.
  • The software remains in beta with expected early-stage stability issues.

The interface unifies all agent operations under a single customizable plugin framework. While current builds experience minor beta friction, the extreme extensibility positions the tool for rapid growth over the subsequent year.

Community Posts

No posts yet. Be the first to write about this video!

Write about this video