Content operations with DeepSeek Harness and Mistral Large 4, aka Le Chonk #
If you’re wondering what it looks like to use agents on websites via Wagtail v8.0’s new API, here’s a demo ✨ There is a lot in there so I’ll explain separately:
- What DeepSeek Harness is and why we picked it
- Why Mistral Large 4 as an LLM.
- How the Wagtail API and skills make this possible.
Content ops with DeepSeek Harness #
The demo goes through three simple content updates on a live production site, our Wagtail user guide. This kind of work can be done with any "cowork" general-purpose agent, but we chose to showcase DeepSeek Harness, which is one of the best options currently available:
-
It works with any AI models / providers, not just DeepSeek.
-
Doesn’t require logging into any account, DeepSeek or otherwise.
-
It’s fully open source, so you can see exactly how it works. And customize it to your needs.
-
It’s compatible with shared standards like Agent Skills So there’s no vendor lock-in whatsoever, unlike all other options. And it has a lot of great features if you care to know how AI models answer your queries, including this amazing "trajectories" viewer so you see exactly how the model uses different tools:
The demo is all based on default / built-in features of this harness.
Le Chonk from Mistral #
It only felt right to demo this with another provider’s model, so we picked Mistral Large 4, a chonky 1T parameter model. We like Mistral because they’re very transparent about the environmental footprint of their models, particularly when it comes to training. We expect the footprint of Le Chonk to be relatively similar or lower than numbers they’ve shared in the past for Large 2, which was a smaller model but had an architecture that was much less efficient:
In this demo, we went through 1M tokens (162k input, 12k output) over 41 LLM requests, for a cost of $0.20.
Making it possible: Wagtail skills and CLI #
So we have our harness and models, but we still need them to be more aware of Wagtail. This is where our new write API comes into the picture, so you can do all sorts of admin operations from outside the CMS. But for the agent to more easily use the API, it’s also equipped with:
- [Our prototype CLI](https://wagtail.org/blog/prototyping-a-cli-for-wagtail/) , so the agent has a more reliable way to execute API requests.
- [Our new agent skills](https://wagtail.org/blog/our-agent-skills-for-wagtail-developers/) , so it’s very clear to the agent exactly how the API works.
We’ve also been experimenting with MCP, to achieve similar results. It’s very promising but certainly not 'needed'. Compared to MCP, command line tools take more effort to install but have a big advantage of being easier to combine together with one-another.
Try it out with your agents #
All of this is meant to work well with any agent/harness, and also be highly suitable for other kinds of automations than AI agents. We’d love to get more feedback from early adopters, let us know how it goes!
And treat it as a teaser of sorts for Wagtail Space 2026, where we’ll demo even more complex content operations with similar tools and models.