正在加载视频...

视频加载失败

I discovered the wildest feature of GitHub Copilot CLI by complete accident last night. You can have models call each other. So you can have Opus, Codex and Gemini all working together in the same chat.🤯

95,048 次观看 • 9 个月前 •via X (Twitter)

46 条评论

joel ⛈️ 的头像
joel ⛈️9 个月前

@github i use claude to run codex because i hate interacting with 5.2 on a personal level 😂

mithun shanbhag 的头像
mithun shanbhag9 个月前

@github Does this count as one premium request? Or multiple?

Burke Holland 的头像
Burke Holland9 个月前

@github GREAT question. I have no idea. @pierceboggan do you know how the billing works here?

Nick 的头像
Nick9 个月前

@github That is really cool indeed, but isn’t this how agents are designed? For each agent definition you can specify the model explicitly. Or do you mean it spawned the agents ad-hoc with the requested models without being specified up front in your config?

Hesu Krypto 的头像
Hesu Krypto9 个月前

@github Interesting... So the model has to be Auto? Or not necessarily?

Burke Holland 的头像
Burke Holland9 个月前

@github No! It can be any model! That said some models are gonna suck at following your instructions *cough* gpt-5-mini *cough* and might not actually call the subagents correctly.

Simon C 的头像
Simon C9 个月前

@github Do you know what I saw this a while back from the amazing @_Evan_Boyle but then completely forgot about it - so thanks again for the reminder! I need to definitely use this more!!!

Erwin Kroon 的头像
Erwin Kroon9 个月前

@github I use it to analyze the content from a screenshot or photo from opus in a haiku sub agent, saves a lot of context, and a bit faster.

Burke Holland 的头像
Burke Holland9 个月前

@github Nice flow - I'll have to try it

Hugues 的头像
Hugues9 个月前

@github I have been using that workflow for a few weeks now (since it released). I no longer use VS Code chat and I'm using 100% Copilot CLI, it's just better. I just switched from VS Code Insiders to the blue logo, since all the updates come from Copilot CLI now :)

Burke Holland 的头像
Burke Holland9 个月前

@github You can do this in @code now too!

Guillaume Berthier 的头像
Guillaume Berthier9 个月前

@github It’s over , It’s over we are cooked. The beast is out of the cage and there’s no hitting pause now. 💀🔥

Franz Hemmer 的头像
Franz Hemmer9 个月前

@github So cool! I think we have the great Steve Sanderson to thank for that, he demoed running a prompt on 3 models simultaneously last week. This is an even better use-case for that capability in Copilot CLI - now we need that in Copilot Chat in VSCode assuming we don't have that yet.

Rob Roy Hobbs 的头像
Rob Roy Hobbs7 个月前

@github where is gemini pro? Seems missing for some but not others...@pierceboggan

Mustafa Ergisi 的头像
Mustafa Ergisi9 个月前

@github Multi-model orchestration is a game changer. Each model has different strengths, being able to route tasks to the right one automatically is the future of AI tooling.

Larry Velez 的头像
Larry Velez9 个月前

@github Can OpenCode do this, @thdxr ?

Shannon Deminick 的头像
Shannon Deminick9 个月前

@github Each skill defined can be tied to a model and in your prompt you can also tell it to use a certain skill for a certain part, then it will sub agent the call - which doesn't cost extra tokens, only the first premium token is counted.

César Couto 的头像
César Couto9 个月前

@github This is the way :)

Aaron Glover 的头像
Aaron Glover9 个月前

@github I wish when you create custom agents, like " the model would automatically switch to the custom agent rather having to say "using csharp-expert agent do X and y".. It even ignores top level instructions like, "when writing c# use the csharp-expert agent"

Josh Ndemenge 的头像
Josh Ndemenge9 个月前

@github Works in the cli 🤯 but not on vs code

Burke Holland 的头像
Burke Holland9 个月前

@github I think it will work in @code though cc: @pierceboggan

Jeet Majumdar 的头像
Jeet Majumdar9 个月前

@github Is it running subAgents with the specified models you told it to. Its kinda crazy,

ashith 的头像
ashith9 个月前

@github Are you sure its not a hallucination?

Burke Holland 的头像
Burke Holland9 个月前

@github The thought did cross my mind, but I doubled checked with the team and it is not.

ashith 的头像
ashith9 个月前

@github Wow thats amazing! Did it call the other models through a tool call?

Burke Holland 的头像
Burke Holland9 个月前

@github Yes - it appears like it's a tool call

Sam Lin 的头像
Sam Lin9 个月前

@github It's a bug. Wait, it just became a feature

Scott Lahteine 的头像
Scott Lahteine9 个月前

@github Good to know! I had assumed this would be a standard feature of AI assistants, if not already. I wouldn’t really want to use an AI assistant tool that forces everything to go through the same model, when different models have different strengths.

L 的头像
L9 个月前

@github does each sub agent cost each their own usage?

neozoid 的头像
neozoid9 个月前

@github …🦝

Jerry 🖖 的头像
Jerry 🖖9 个月前

@github I do this manually today.

Sandip Bhattacharya 🌏✌️ 的头像
Sandip Bhattacharya 🌏✌️9 个月前

@github 🤯This is an incredible advantage for multi-model harnesses over single-model harnesses!

Burke Holland 的头像
Burke Holland9 个月前

@github RIGHT?!?! Why on earth would you lock into one model?

Brad Germain 的头像
Brad Germain9 个月前

@github What if I use the cheaper model to call the more expensive model to do the work? 🤔

Richard Shade 的头像
Richard Shade9 个月前

@github Have you tried it with code-review or speckit? I could see multiple models double checking the review, plan or tasks phase. I have actually been swapping a bit to gemini for `/speckit.plan` as it has such a bigger context windown and web_searches easier

vincentb 的头像
vincentb9 个月前

@github Make a skill in claude to call codex (or any other cli coding agent). Only one tool needed, leverages many models as you wish

Gregor 的头像
Gregor9 个月前

@github Sometimes these accidental discoveries lead to the most innovative uses. I've stumbled upon similar unexpected features while building my app, it's amazing what you can find when you're not looking.

Tarun 的头像
Tarun9 个月前

@github Bravo, thanks for sharing!

Jatin Puri 的头像
Jatin Puri9 个月前

@github Super awesome to see this. I wasn't aware either!

neozoid 的头像
neozoid9 个月前

@github WOOOCOOL!🔀

foundrceo 的头像
foundrceo9 个月前

@github this is wild man, collaboration just got a whole new level fr

V 的头像
V9 个月前

@github Interesting, must try

Burke Holland 的头像
Burke Holland9 个月前

@github I can't believe this isn't a bigger deal. This seems like a massive unlock, but I'm not 100% sure what workflows it unlocks. Which is like, just another day with AI.

Treppenwitz 的头像
Treppenwitz9 个月前

Wow. I am not a coder but have managed to build a site and am using copilot with Opus, Gemini, ChatGPT Codex, and a few others. I’ve been looking into extensions to use them as a team/swarm to fix bugs and build faster. If this can be done my adding the directions at the end of the prompt, that would be amazing. Is that all you are doing?

Burke Holland 的头像
Burke Holland9 个月前

@hirletz @github That's all I did here. But you could create custom agents that do this. Agents that call agents that call agents, etc. Basically orchestration. Something like

Faheem | FrontierMind AI 的头像
Faheem | FrontierMind AI9 个月前

@github Multi-model chats are fun until coordination debt hits: shared context, conflicting tool calls, unclear source-of-truth. Do you run a single orchestrator + specialist workers (with state), or let them free-for-all?

相关视频