Wagtail API in action: Content ops with DeepSeek Harness + Mistral Large 4

Hello! Today I thought it would be cool to showcase how lots of things we've been working on in Wagtail come together to help with the content work on the sites. I thought it would be nice to do it as a live demo of the capabilities rather than talking about them in theory. So we're now inside the agent harness called Deep Seek Harness. It's pretty new, it's from DeepSeek, you guessed it. But despite the name, it works with all types of models out there, definitely not just DeepSeek ones And all types of providers as well, so it's not restricted to your AI inference running on DeepSig servers And just to demonstrate that, I thought I'd combine it with the latest uh Mistral

Large 4 model nicknamed Le Chonc. Love that. just to showcase that yes in practice you can have an EU-hosted model running on like all sorts of content management side tasks on on your site So just to be clear, like all of this we're looking at on the screen here is Deep Seek harness. It's just a it has an embedded browser that I'm using to showcase. this. And yeah, I'll demo this on the Wactile User Guide website, which is a relatively simple site as far as the control model. But what I'm demoing today is relevant for all types. of sites. Um so I think I'll just like start typing, write a prompt and we can see how it works. Hey could you help me fetch a copy?

Of the latest manage sites page via the ArcTail API. Um so pretty generic prompt there. There's some indication to the agent of how to do this. There's two specific keywords here wagtail because I would really like to make sure that this loads the relevant skill that the agent is aware of to do work on wagtail websites. and then API because I wanted to load information about the API specifically. So we're working lots lately on making sure that agents via our new CLI the prototype of interacting with the CMS via the command line, have access to the right information as well as the right tools.

So here this is just to see whether it can find the information. This is a pretty new page I've worked on relatively recently, this uh manage sites, and I think it's just a good demo of um how much the API can help with tasks like this uh with the CLI or otherwise so while we're looking at this I'll head over to the the the CMS in parallel and uh I have a specific thing I want to do with that page that page is already live but currently it's not visible anywhere because it's not set to show in menus so I want to make sure to ask uh the agent to help with that uh so I'll just assume that he's successfully fetched it at some point and write my uh task accordingly now

could you please set it to show in menus okay It's gonna queue that so that as soon as the agent has finished with the previous task, it asks for this. next and um the next task i want to do after is um on this page there is um need i think for a link to a separate page we've added recently on the site So if I navigate in the CMS, you see I worked on managed sites but also on this new multi-site page. And it makes tons of sense to make sure that

the multi-site page is linked from managed sites just like managed sites is linked from multi-site so having it go both ways um so oh and please also sets the multi -site page to show um so here you know i'm doing it as two separate tasks once per page but i could also do it in bulk You'll have seen you know these live updates, this this kind of canvas interface that shows like files having changed here. So those are all temporary sites. There's not necessarily any need for me to retain them. It's just inferring from the environment I'm running this in what to do.

I really like Deep Seek harness compared to Cloud Coalc or Codex because it's well first it's support any AI provider, not just uh big ones. Um I can have it set up with as many providers as I'd like at the same time. So here I have three right now deep seek itself tensor eggs which we use lots for wagtail stuff and uh and mistral but i could have as much as i want It's also very transparent about the total usage of AI inference, as well as coming with lots of tools to inspect exactly how the AI answers my queries, which you know, since I care a lot about making sure that Everything is as efficient as possible. Getting this kind of information on like you know how many tools the AI interacts with, what's successful and not, that's uh that's a big difference.

Okay, so it says it's done the tasks, gives me the link to it. I think what we can do quickly is just check in the browser whether that is uh done or not. So sorry, I'll just just uh yeah make sure to switch to the right page go to the live version and I should see it in uh the main menu of the site yeah so manage sites it's available there now and in reference we now have multi-site so that's all done correctly and now ask for a content update. Hey um see how multi -site links to manage sites could

you make content updates to um the Manage sites page so it links back to the sites. So obviously as someone who's taken the time to create those pages, I do have like a pretty good idea of what kind of linking I'm after here uh but I think just for the purpose of the demo it's really interesting to see how it executes those tasks where it might decide to create those links. If I head over back to the page content, it'll probably be about having reference information in the about area here.

or maybe at the bottom as kind of like further information link. So this is something we're still working on, like making sure that uh agents making those kinds of updates have as much awareness as um possible about the site's content model We have this Wagtail space event coming up in a few weeks that I will definitely be able to reveal some of those latest changes at. It would be very interesting to look at. and uh will also feature how uh organizations out there have been using this API uh so far with or without their agents. Um okay so it looks like it's made some progress so we'll head over back to the CMS.

And um there's definitely some ambiguity here on whether it's making those updates right away to the published version, like I've asked before. or whether it's doing it as a draft. So just to be clear, by default the API access has the same permissions as your own user account in the CMS. But you can definitely customize that if you want for it to have less access and mandate that there always has to be drafts for you to then manually look at. Okay, so it's made the changes. I think um in the about sites section and that makes sense to me so here I can see that I have a conflict with another update so be able to click that and see what updates he's made. Yeah, so right here.

Most projects have a single site, but Water also supports multi-site installations. Can switch back over to the previous version that I already had opened to compare. Could also look at the built-in compare view in Wagtail, but that's just not worth it for such a small update. Yeah so here I think just add objects to it taking so much space. I feel like one link to multi-site is plenty enough and um oh okay so it's a separate link to the developer docs and to this website's docs so that's probably that's probably okay then um what i definitely want to make sure of as well is that it links as an internal link

I think it does, but never 100% sure at this point in time with RUX we need to improve that. And um Yeah, I think if that's not already live, I'll definitely make sure to publish this version. Um so yeah, back to back to the harness. Um really there's no need for something like MCP in place on the site here just by having this little command line tool available on the computer, it's already able. The agent is already able to make all of those updates operations. We could definitely do something like this with MCP. in the future but there's really no it's not a blocker for us to be able to access those kinds of capabilities um and yeah just uh again

as someone who really cares about exactly how this uh is operated by the agents having access to this kind of information tremendously helpful to know or perhaps we could make sure that a content update like this you can do it in like say four tool calls rather than eight or perhaps the initial discovery of the API we should try and optimize that so that happens in like 10 tool calls that's just invaluable and um yeah just as far as cost management uh having a clear idea of you know what we're paying for exactly here is it just the harness itself having very like uh expensive things to load is it just the conversations we're having

uh super important And yeah, again, I do feel like it's a big deal that we can run AIs with providers of our choosing. So in this case, uh Mistro, they're one of the most transparents as far as the environmental footprints of uh their AI training. This is now part of the law in the EU that they have to share those numbers, but Mistrol did that well ahead of its having been. part of any kind of legal framework. So they're really good at this. And um yeah just you know it's very um one one less thing to worry about for us to know that that uh the AI usage, not just the training, is done uh in a country like France where uh we know that the electricity grid is uh quite low carbon.

So yeah, putting it all together essentially it means that we'd we'd love, I think, people probably able to use that tools like that to work on the Aracted websites if it's a good fits for for those sites for your initial needs. It's definitely possible. I think the only question really and comes back to you is at which pace do you want to adopt those those things if at all and that's all very configurable and um yeah I hope we'll get to showcase more examples of people using those tools helpfully at uh Wagtail space come uh November Right, uh well until the next one