Inside Project HydraFusion: multi-model orchestration in the GitHub Copilot | demo | GitHub Checkout

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Inside Project HydraFusion: multi-model orchestration in the GitHub Copilot | demo | GitHub Checkout · Channel: GitHub · 13:04 · Watch on YouTube

Topics: AI agents, AI developer tools, Model inference and optimization

Summary

Project HydraFusion is an experimental GitHub Copilot CLI research preview that selects an execution strategy based on the user’s task, so developers can focus on their prompt instead of choosing models. It uses three patterns—single, cascade, and critique—and can involve multiple models when a task needs escalation or review. The demo compares HydraFusion with other models and highlights token efficiency, while noting that the critique pattern can use more resources for complex tasks. The team is working on broader availability, improved transparency, and a future public preview, and encourages users to try it and share feedback.

Key points

HydraFusion is available as a research preview in the CLI after enabling experimental mode and selecting it from the model list.
Unlike the earlier capacity-driven version of auto selection, Hydra considers the user’s intent when choosing how to handle a task.
The single pattern assigns a task to one model for straightforward execution.
The cascade pattern starts with one model and escalates to a stronger model when needed.
The critique pattern uses a back-and-forth between models to draft, review, and revise a plan for more complex tasks.
The demo showed HydraFusion using fewer AI credits than the compared models for several tasks, though the critique pattern can require more resources.
HydraFusion displays the selected execution pattern and AI credits, but the team is considering how much information about the specific models to reveal without biasing users.
Planned work includes bringing HydraFusion to the GitHub Copilot app and VS Code, improving the user experience, and moving toward public preview.
Users can try the research preview through the CLI and provide feedback in the external discussion post.

Chapters

Start	Chapter	Summary
00:00	Introducing Project HydraFusion	The hosts introduce HydraFusion and explain how to enable its experimental CLI research preview.
01:33	How HydraFusion chooses an execution pattern	Julia contrasts intent-based routing with earlier capacity-driven selection and describes the single, cascade, and critique patterns.
05:12	Demo: patterns and token efficiency	The demo compares HydraFusion with other models across single, cascade, and critique tasks and discusses their AI credit use.
08:22	Visibility and model-selection transparency	The speakers discuss which details HydraFusion shows and why revealing specific models could bias users.
10:29	What’s next and how to try it	Julia outlines plans for the Copilot app, VS Code, UX improvements, and public preview, then invites users to test the CLI preview and share feedback.
12:33	Closing	The host recaps HydraFusion’s three patterns, directs viewers to the discussion post, and closes the episode.

Quotes

You can essentially just use this as any other model where you just give it a task — 01:02

One is called single, one is called cascade and one is called critique. — 03:39

Hydro fusion takes care of the rest. — 05:12

we do like transparency as well, right? — 09:57

One prompt end and it picks the pattern you need. — 12:33

Concepts and entities

Project HydraFusion, GitHub Copilot CLI, GitHub Copilot, VS Code, single, cascade, critique, Julia, Opus 5, Fable 5.1, Luna, Soul

Full transcript (original language)

00:00 what’s the latest model I should be using? How should I be using it? They don’t have to worry about what model to pick. You can essentially just focus on the task you want to execute. Julia, welcome to GitHub Checkout. Thank you so much for having me. Super excited to be here. >> I am so excited to talk to you today because I have long wondered how these things come to be and what I say these things. The project we’re going to be talking about today is considered a model. I guess it will be like our first dip into that technology, but it’s also

00:31 a technology that we’ve been using with GitHub copilot uh to help our users make better model selections and be more efficient when they’re working with Copilot and that is you’ll have to explain to us what project hydroofusion is, how it came to be. Let’s start from the beginning. >> Right now it’s available in the CLI. It’s under experimental. So make sure whenever you’re testing it, you’re first doing SAS slash experimental. And once you’ve enabled this, you can just do

01:02 slashmodels. And now you’re able to find hydrofusion research preview in your model selection. And it’s very easy. You can essentially just use this as any other model where you just give it a task, any kind of prompt that you have. It will then go and start executing it but it will pick the right pattern for you because you don’t essentially have to worry about it how difficult you task it any of these kind of things and then it will go and execute and start implementing for you.

01:33 >> Could you tell us a little bit about the difference between this and auto because we have been using auto model select as well. Is this the underlying technology for auto model select or this is something completely mobile? So originally I think at the beginning of this year we introduced what we consider auto version one and it was very much skew or capacity driven. So a user was typing in a text prompt and based on we knowing where the best availability the best capacity was. We were routing to

02:05 that model in that moment that sounded or was the perfect fit. But we saw that capacity routing isn’t it’s one of the things that people are considering but it’s it might not be the perfect one because at the end of the day it is depending on your intent. So that’s where in May 2026, so this year we released within auto what we are calling hydra we already looking at the intent and rather than just routing to the best capacitywide model we are scoring your

02:38 intent on four different ways to score it and depending on the highest score that’s where the optimization problem is already showing. Hydrofusion is essentially taking it to the next level rather than only you picking one model and that’s it. Now that’s where we are introducing cascade critique where it’s going to be more of a back and forth between these models and that’s where we also seeing a lot of these um efficiency or token efficiency gains overall because we are not just relying on one model. We are relying on a bunch of

03:09 different models and depending on your task we can pick the right execution strategy for you. For a lot of us, we’ve been doing this manually, right? Like we’ve been sort of using this model for this because we have the impression that they will execute better for that task. But not all task are built equal. So if you could talk us through what that decision making that leads into the pattern selection works like I love to understand that better. >> So there are three ways how this could be done. One is called single, one is

03:39 called cascade and one is called critique. single one is the most um straightforward way how hydro fusion will go and solve your task. Think of you typing something in and it will pick one model and it will go and execute and great here’s the answer and that it has chosen and we have the confidence that this is the most token efficient or efficient answer but also the most correct one. Then we have the two other one which we know sometimes tasks can be more complex and this is where we are

04:10 either delegating to cascade or either delegating to critique. Cascade you can think of okay you’re giving one model this task that you have intended it’s trying it with its first um past and great if the answer is correct and we have a lot of confidence that this is this is looking good perfect this is the way where we will go but if you’re if you’re not there yet and if we are thinking hey there is another path that is needed we will be essentially escalating it to a stronger model so we

04:41 are using another model to come in and it will help it go and execute it and perfect. It will come up with its own answer. And then the third one which is the most complex one which is considered or called critique. And here we will have a little bit of a back and forth between models where we have an author model and we have a critique model and they are helping each other to come up with a drafted plan before going into the execution and into the answer algorithm and trying to give our end users the perfect and great answer. And

05:12 the interesting thing is we’ve seen these kind of patterns users or developers already been doing for like for years once we started introducing new and new models >> on the user. You don’t have to select anything else. Just setting out your prompt and that’s it. Hydro fusion takes care of the rest. >> That is correct. So I even have a short little demo with me. All the way on the right you can see we picked Hydra Fusion for our intent and um we are comparing it across Fable 5.1 and Opus and as you

05:46 can see Hydra picked single as the execution um strategy. I just wanted you to know that all of these are the same task. So we are essentially giving all of these three different models the same task, the same prompt and we we were curious on what it goes and does on the level of token efficiency. So that’s where you can see all the way down here what are the AI credits that are being used as another example. So that’s why we are doing it again. We are giving it another task. Again it picks single here

06:17 and it picked single as the right execution strategy. And again you can see the AI credits being used across Opus Fable Fable 5.1 and Hydro Fusion. And again hydrofusion here is one of the winners in ter terms of token efficiency. to demo the cascade execution pattern. First a model was picked. It then needed to escalate it to a more complex one. So here for drafting it used Luna. Then it used soul for checking and then it escalated it again

06:47 to soul because we needed a little bit more power than Luna can were able to provide us. And at the end of the day hydrofusion is still the most token efficient one if you look across opus and just fable 5.1. And then the last execution pattern we already touched on which is called critiqued. This definitely has the highest fast radius. It’s one of the most complex um execution strategy that could be picked because here we have Opus 5 as the model that is using the drafting. It then uses

07:20 another very strong model which is 5.6 soul to review it. And then all the way the last step is where it’s revising the plan and here it’s using opus 5 again. So there’s a little bit of a back and forth and the reason behind that is because we had detected that this is a very complex task. So we needed a little bit more power overall or if you think of the total Hydro Fusion is definitely showing the very exciting token efficiency gains that we were having knowing that with the last one we needed

07:51 um to make uh the decision for picking critique um for going back and forth between the two very strong models. But we had so much confident in picking those two stronger models that we are we were willing to take the little bit of higher token um efficiency um credits here into consideration and to make sure that the answer our end users are getting is the right one that they want. But overall, if you think about all of the different tasks you can

08:22 be doing, hydrofusion is still one of the ones that is showing um the very high token efficiency gain. >> Once the attacks are executed, like do you get to see what pattern was selected by hydro fusion? >> So two things. Yes, we are already showing what pattern has been picked. And if you remember my demo all the way at the beginning, we were able to see that single was the one that was picked. So we are showing what’s the pattern we wanted to show. And also you can see at the end the AI credits that were used for it. The things that we are not yet

08:55 showing and that’s something we are debating right now is do we want to show people what models have been used and the reason behind that is and that’s why we originally didn’t go out with showing was we’ve done some experimentation and if we are putting a certain model in front of our end users there might be a bias whatever we have picked as the model we have these very high level um thoughts about the latest models, right? Fifi one, Opus 5. We know these are very

09:27 strong ones, but also very costly ones. So, we didn’t want to bias users by maybe seeing us picking Luna, which is still a very capable model, but we knew in that moment for drafting your plan, Luna is totally um okay for using it. And we didn’t want to bias people seeing a specific model and having certain thoughts. So, that’s why we made the design decision to not showing it. We do know we have to bring in some more information for our end users because we

09:57 do like transparency as well, right? Yes. So we have to find the right balance between what we want to show without putting certain biases towards our end users. >> I appreciate that. At the end of the day, the best part is that that top process and all of that complexity you don’t have to worry about as a developer. You just focus on the task ahead. keep building and let co-pilot and hydrofusion do the thinking and all these very layer very complex decisions that are happening underneath can do it for you. So I love that. So what’s

10:29 coming next like what’s in the plans for uh the project? First of all I always recommend check out our blog. It has a lot of more details if this is something that you are interested in. It also explains on how we benchmarked right before our release and we do have an external discussion post. So I always encourage people to go in and respond there whatever you’re feeling especially while we are iterating on it. So what are the very two things that we are working on right now? The very big one is um for adoption but also get it into

10:59 more hands of our developers who don’t necessarily just use the CLI. So that’s why we want to make it available in the GitHub copilot app but also in VS code. So this is being worked on right now. And then also we already talked about UX improvements, right? We have to find a right balance between what we want to shout show to our end users. We want to provide you all transparency and visibility and what is happening while also making sure we are not putting these biases in or things that we have learned from experimentation that we

11:30 don’t necessarily want to show to our end users to put any kind of bias into their head. But we know there’s definitely some things we have to work on. The next step is public preview which is also super exciting. We don’t know yet when, but it’s definitely already on top of our minds. >> I appreciate that. And listen, this is why we need you um our users to go and enable it. So, all you have to do is to go slashexperimental on. You have to turn experimental mode on on the CLI to be able to enable it and then

12:01 slashmodel, which is going to bring you to the normal list of models, but you’ll be able to see project hydroofusion there. >> Whatever you’re working on right now, these are the tasks we want at the end of the day. We want to build features, functionalities for your daily um drivers, for your daily tasks. So just go and try it and let us know how it’s working for you especially around token efficiency as this is one we’ve seen in our offline evas but also while doing a lot of online experiment and online testing me myself the entire research team. So no biases here just go have fun

12:33 with it and let us know um your thoughts on it. >> Amazing. Thank you so much, Julia, for bringing project hydroofusion to us. I can’t wait to see what’s next. And that was your first look at project hydroofusion. One prompt end and it picks the pattern you need. Single cascade or critique depending on the task at hand. If you want to try it, turn on experimental mode in the CLI and run /model to find the research preview. And please don’t forget to leave your feedback. Julia and the team are reading

13:04 the external discussion post every day. So if you had something interesting, please tell them there. And don’t forget to subscribe to this channel so you never miss another release or feature update. Push the changes domain and I’ll catch you on the next release.