What AIs are currently relevant for Unity?

[thin hill]

There's a lot of information on AI, but it's not organized into a manual. Some of it is definitely outdated. I'm a newbie and don't understand a lot of it. So, I'm asking which AIs to try with Unity 6 to start.

[desert dune]

What exactly are you looking for? What kind of AI? For what purpose?

[thin hill]

I don't have a precise answer. I'm looking for all the AIs that I can work with and that are relevant now and in the near future.

[desert dune]

You'll need to be a bit more specific. Are you looking for a chat bot to answer questions about unity? A coding assistant/agent? AI for in game features? Asset generation?

[thin hill]

I'm interested in all AIs that can help. I have no knowledge of how AI and Unity interact.

[mortal eagle]

You can start here
https://unity.com/products/ai

[thin hill]

Is this AI alone sufficient for the job? Or are other AIs possible? If so, please provide examples.

[mortal eagle]

What job? You never mentioned what specifically you need to do. Most tools available are fairly specific in doing one thing. There is no one AI that can do it all

[thin hill]

If it's not too much trouble, could you please list the AIs that can be used?

[mortal eagle]

Can be used for what? Literally all them with a public API can be used. Others can be used locally by running the ONNX through Inference Engine.

[brave basalt]

there is bezi that is not bad, it can help you with your unity project

[thin hill]

So Unity users use any kind of AI? There's no current list of AIs?

[mortal eagle]

There are hundreds of AI models out there
https://openrouter.ai/models

[desert dune]

Most people that do use, use coding assistants/agents, like those provided by github copilot, open Ai codex, etc...
Some use chat gpt or other chat bots for general questions and help.
Some use it to generate assets(no clue what's available and common these days).

None of these I would recommend to a beginner. You need to be able to do what the AI would do for you, or you wouldn't be able to spot issues/mistakes and fuck up your project.

[desert dune]

And no, there's no list of "what AI people use in unity". You're free to compile your own list if you want, though I imagine that would be a pretty difficult research.

[thin hill]

This is truly a complex study. That's why I asked for help. To avoid wasting time searching for what other developers have already found.

[ionic jetty]

I wanna know to

[chrome apex]

You could ask an AI. They are all great at answering vague questions.

[thin hill]

AIs respond unfairly to other AIs. Their responses are often biased. For example, Deepseek often ranks near the bottom of the AI ​​response rankings.

[grim sentinel]

I like Bezi, its a plugin you can install into unity and it modifies your files for you, you can just prompt, make a script that visualizes tic tac toe, and you got it

[winged oracle]

can you genrate 2d sprite animations using unity AI?

[sterile rock]

I cant find unity AI I'm on 6.2 6000.2.9f1 it's enabled on unity website and for my project but it still no work.

If someone knows how to find it please let me know

Clarify: I dont know if it works. I cant find it. It's not there

[desert dune]

Quick google:
https://docs.unity3d.com/6000.2/Documentation/Manual/ai-menu-access.html

[mortal eagle]

Not directly. Might have some luck using an existing sprite sheet as a reference image. The Animation Generator is only for 3d animations

[winged oracle]

how those online software like voidless or pixel lab animates character from an input?

example

[desert dune]

Probably trained(or fine tuned) an image generation model on inputs.

Assuming it's really generated by a model.

[mortal eagle]

If I had to guess, I would say they are injecting pose references into the prompt.

[winged oracle]

hard to achieve or train something like that?

there are so many online cc0 animations to get the data

i was thinking using OPen pose control net apply but i do not know how they retain the same chaarcter

[desert dune]

Need a lot of training data. Though if you're OK with overfitting, little data would work too.

[winged oracle]

well characters can e generated using AI fro the image input

[desert dune]

Pose references might be easier, since you can leverage existing models, like qwen image.

No extra training needed.

[winged oracle]

i am just thinking how to replicate this that is all

[desert dune]

Well, we gave you several options.

[winged oracle]

going to try that out using pose references

[sacred rose]

Using https://aiplitu.de for my unity questions

[uneven skiff]

I am using Bezi

[upbeat timber]

if you need one to answer programming questions, claude is better than chatgpt on both free and paid tiers. i havent looked into any other AI yet.

[floral bobcat]

I'm just starting to get into AI in unity dev'ing and so gonna bump this thread

Just a start. What's the news with unity sentis?
Anyone have tried doing the things mentioned here?
https://discussions.unity.com/t/bringing-runtime-ai-to-unity-6/1529799

I'm like 3 years being out of the loop with these AI stuff, and gamedev in general

Bezi looks cool that it integrates inside unity editor. But it's not free.. hahah

https://www.youtube.com/watch?v=9sNKulZ1-ho
Currently watching this

Basically i'm willing to pay, maybe $20/month at most, for the right tool. So right now just doing all the research i can

[desert dune]

Well, what exactly are you looking for? These are vastly different tools for different purposes.

[mortal eagle]

If you want to run models locally then Sentis is what you want. HuggingFace updates the Unity models fairly regularly. Inference Engine got renamed back to Sentis recently

[floral bobcat]

Oh it's called sentis again? Haha ok

[floral bobcat]

The most important is definitely a coding assistant that knows codebase context

I know the option, but just looking for the cheapest i guess

And i wonder if any of the coding assist (or unity debug assistant in general) can figure out some obscure unity issues

[floral bobcat]

Like this one.. i'm seriously at loss

[desert dune]

Well, then basically codex, github copilot, claude cli or gemini cli. Or the corresponding vs code extensions.

[desert dune]

They might be able. It depends.

[floral bobcat]

I fixed it by restarting unity

[floral bobcat]

Like, this problem shouldn't even happen. Attempting to "solve" it will make me waste time, especially if i just follow whatever AI says
Restarting unity is an experience thing
Sometimes i think unity editor can trip out worse than any AI hallucinations

But anyways that's irrelevant to the topic

[floral bobcat]

How are they better than unity's builtin AI?

And a relevant, general question. How can an AI know if their suggestions are correct/effective? Like are they able to test by playing, see the game view, notice rendering issues, etc?

[mortal eagle]

Unity AI Assistant should do all of that as well. I believe it's still free while in beta

[mortal eagle]

Generally it can not. It's why AI gets crapped on so often

[floral bobcat]

If my question is like, what's best to approach lighting for my specific project. There's typically indoor and outdoor parts in a scene, with day/night
For indoor, baked lightmap or realtime?
For outdoor, baked lightmap or realtime?

If the AI suggests some things, they still have to get consecutive feedbacks from, i guess me, to see if their work is correct?
Are they able to.. like if i feed them (or maybe they can see themself) the rendering pass in the game/scene tab, can they interpret it correctly?

[mortal eagle]

AI is the absolute worst for subjective questions like this.

[floral bobcat]

If i use some unity asset, and the dev just aren't responsive, would the AI (i guess i tell it to read the docs) know how to work out that asset, with another asset?

[mortal eagle]

It might be able to tell you which ones perform better

[floral bobcat]

Generally when i say "i'm looking for an AI assistant that can know context", this also means the context of some unity assets too
I guess the problem of "the asset dev are not responsive and no one can help me" is still something AI can't do much about

[mortal eagle]

This would likely depend on how popular the asset is

[floral bobcat]

I guess i'll be specific and mention, Bakery, unity's APV, Magic light probe/lightmap switcher (in order of popularity imo)

I guess... from ur answers, AI usecase is still not the best for larger codebase (in this case unity asset stuff)

[mortal eagle]

If the code is in the package, AI might be able to reason about it. It's a compiled dll probably not

[desert dune]

I'm pretty sure what happened is that unity modified the project settings or some other file, but perhaps in memory state remained. Something like this would be clear if you were using version control. Also, most ai agents would probably notice that.

[desert dune]

Just from my limited experience with it, it's way weaker model, and the tooling is also not as sophisticated as the solutions I suggested.

[desert dune]

There are unity MCPs that provide the agents with tools to inspect various project aspects, scenes, prefabs, settings and modify them.
Without these tools, the best it can do is read the project files and modify them.

[desert dune]

If these assets come with source code, then yes, AI agents can help a lot.

[floral bobcat]

Yea but whether the result is visually good is.. anyways i'll just have to try

[floral bobcat]

Btw yes i found where it was from checking the VC. I was looking at the wrong file

[desert dune]

Weren't we talking about code? What does it have to do with visuals?

[floral bobcat]

UniversalRenderPipelineGlobalAsset

[floral bobcat]

Lighting is about visual

[desert dune]

Yeah, but if it's a specific feature implementation that's entirely objective.

[floral bobcat]

Like if i notice some lightmap leaks, i'll.. i guess describe it verbally and hopefully the AI can suggest som things to check. But i wonder if it can check the scene view, figure out which mesh is doing this, check the vertices for some bad form, and figure things out itself

But wait2 anyways this is getting out of topic. I really need to just try things myself

[floral bobcat]

Which one out of these needs the most basic setup?
I'm currently trying github copilot with visual studio 2022 (need to update mine apparently)

I have some bad experience with VS code before and would like to stick with just visual studio if possible

And.. these suggestions are the ones that can uhh.. be aware of the codebase context?
Maybe i'm assuming github copilot needs github or needs a repository at all for it to work. Is that correct? I'm using plastic scm if that matters

If it doesn't need a repository, cool. But if it does, does that mean it'll also be aware of different commits, branches, and can give suggestions like "oh how it was implemented in branch Abc is already correct, so better merge that first" or something?

[desert dune]

You could share a screenshot with it. But I don't think there's any point in "checking the vertices".
It might be able to suggest something, but so would just chat gpt.

[desert dune]

Most of these tools are either CLI based or at best VS code extension. The setup is not that difficult for vsc extensions. For the CLI tools it takes some instruction following.

VS only has support for github copilot afaik. Which might not be useful beyond coding.

And.. these suggestions are the ones that can uhh.. be aware of the codebase context?
Yes, all of these tools would read and search the project for relevant context. Though, it might not be ideal. It depends on your project organization and consistency. Think of it as if you let another person that sees your project for the first time help you.

Maybe i'm assuming github copilot needs github or needs a repository at all for it to work. Is that correct? I'm using plastic scm if that matters
Git can be useful. You can create a local repo. If Plastic has CLI api, Codex and the CLI solutions can use it as well.

[floral bobcat]

Wait so u mentioned github copilot in ur list for coding specific
But lets say i also use unity editor's AI in parallel for the general unity editor stuff, that'd make a pretty good workflow in general, right i guess?

[desert dune]

Not sure. I haven't used unity's AI enough. But from my little testing it was not very useful. I personally use Codex, and sometimes github copilot with Claude or Gemini for planning.

With both having access to unity MCP. Though it's pretty junky and would often fail or take a lot of time to get right. Modifying stuff on the editor side manually is more productive.

[floral bobcat]

Ok github copilot is fun but i ran out of token hahaha

I have a few followup questions but for now, is there anything else than github copilot that has more token limit and free?
Or, just free. So i can use up copilot, then use up the other one, etc

Generally i like chat gpt better but maybe it's bcoz my custom instruction

If need more info, ask more questions about the context

After getting my approval about ur understanding of the codebase, then u may explain the bug, correct code, explanation of correction, etc

I'll post scripts, as part of a bigger codebase
Will also post a question or issue
Do NOT respond with a solution before u have enough info. U MUST ask for more relevant info/code from this codebase, when the class is called etc

Do NOT DARE to post any solution before i confirm all of ur understanding of the codebase first

I guess i can have custom instruction like this for github copilot too

It's just too wasteful when it spits out full code response each time. It also doesn't seem to have a "edit prompt" to jump back to previous part of convo

[chrome apex]

copilot is by far the cheapest access to the actually useful models you can get right now. understand that you need to use copilot in agent mode and to instruct it to make a planning document as output, not code (it should do a planning step internally anway, but you can ask it to make it explicit). the "ask" function in copilot is not what its actually useful for. you should treat agent mode as a pair programming session, code review and exploration tool. guide it with strong value decisions and course corrections, curb its enthusiasm but allow it to find solutions you didn't think of yourself.

free tiers on claude code, cursor and chatGPT/codex are not giving you the full experience. you will have to accept that code agents will cost a bunch to become useful and you need to figure out if that cost is producing enough value for you. wastefulness is still a necessary part of the game. the whole technology is incredibly inefficient at its core.

that said, the straightest path with the least drift and iteration from plan to working solutions is achievable with claude opus right now. but you still need to give it actually solvable problems: narrow modules, clear input/output and an architecture it can actually self-evaluate and reason about.

[desert dune]

There's open Ai Codex(I'm not sure if it's free, but definitely good limits with a chat gpt plus user), Antigravity from Google basically free with good limits(using it a lot lately). Some ClI agents from different companies also have free tiers I think.

[mortal eagle]

I think its also worth noting that the latest preview version Unity AI Assistant now has an Agent Mode not sure which models its using though. There are also a load of new Generators for 3D models and Music as well

[uneven skiff]

I like Bezi too.. I believe they are great in this.. I found even they implemented lots of advanced techs to reduce context bloating and LLM forgetting things. It works in Agent mode as well.. very strong tool!

[atomic walrus]

Do you have any more detail about the available MCP servers you don't mind sharing? Any you've found particulalryl useful?

Quick search turned up https://github.com/CoplayDev/unity-mcp, but I'm wondering if there are also others that are good.

[desert dune]

That's the one I'm using. I think it's the most solid one. I've only seen one other, but it's not updated as often.
Other than that I don't use any mcps currently. I'd also recommended configuring it such that only the tools that you really need are exposed to save on context.

[quartz verge]

how to get access to https://unity.com/features/ai I throw my mail in there and nothing comes to me.

[desert dune]

Did you try opening it in the editor? Perhaps you already have access?

How do I get access to Unity AI?
-
When you download the Unity 6.2 beta, you will see a new “AI” button persistent at the top of the Editor window for easy installation and access to Unity AI. During the beta period, all users who install Unity 6.2 and choose to install the Unity AI features (packages) will receive free Unity Points to try the features. Unity AI will enter GA (general availability) along with Unity 6.2 later this year. The free points will expire at the end of the beta, and paid Unity Points will be required going forward.

[quartz verge]

i have Unity 6.3 and Unity 6

[desert dune]

That's not what I asked.

[quartz verge]

I'm asking if there's an AI on these versions.

[mortal eagle]

Its in Unity 6.3+

[vital imp]

For now, we also supported 6.0, make sure you are on Unity 6000.0.60f1

[floral bobcat]

So i've been mostly still using chat gpt...
And for my use i needed to work on several scripts. I ask it to put each script in different canvases
But apparently it can't "see" all the canvas, nor can it edit multiple canvases..
It'd sometimes just make new canvas for a newer version of the same script

Asking it to put everything in 1 canvas works very nicely. But this last new chat session i'm asking it to do that (also with the same 5 scripts) ends up saying "canvas reaching limit"
Idk what else to do with chat gpt anymore. I'm only using free ones for now
I guess will try github copilot again

[desert dune]

You could get some decent coding agents for free. Gemini cli has some free plan quota. Something like opencode provides free access to some decent open source models(better than what github copilot provides for free IMHO).

Also, open router usually has some free endpoints to relatively decent models. You can use it's api key in opencode and maybe github copilot.

[floral bobcat]

I'm otw to try gemini cli free

I still don't get if i have to use github for this or that's optional. What if i'm working on a non repo project?
Or what if it's in plastic scm?

I'm doing this one Gemini Code Assist on GitHub

Or that's different

https://developers.google.com/gemini-code-assist/resources/quotas#quotas-for-agent-mode-gemini-cli
Gonna ignore that one

[desert dune]

Gemini cli isn't related to github in any way. You're probably confusing with github copilit(which ironically should work without a github repo as well). Version control isn't really related to agentic coding tools. Though, I'd advice always using a local git repo for managing/keeping track of the AI changes properly.

[floral bobcat]

Ya there's just an option for the github
https://developers.google.com/gemini-code-assist/docs/review-github-code

But anyways

I'm installing gemini cli and when i authenticate with my google account, it says i now also have access to antigravity

Which is another interesting thing i'm curious about. Anyone tried antigravity for unity stuff?

It's an IDE right

[desert dune]

Yep. The concept is not bad, but the issue I had with it is that the C#/unity intellisense basically doesn't work, because the required packages don't install outside actual vs code.

[floral bobcat]

Well that kinda sucks

Hopefully they include that soon

But.. (i never used cli AI before) isn't that pretty much worse in cli coding then?

This gemini cli and google antigravity are both assuming if u wanna go for agentic vibe coding, right? So.. minimum manual coding or even reading the codes, just checking result once in a while

And i guess, for me, if i still wanna check n edit code once in a while (just comment out, try manual values), maybe i open the same scripts in antigravity and also in VS, is possible right? Just means gotta reload the script again in both IDE, right?

[desert dune]

They won't. It's the extension side issue. Microsoft(that maintains the extension) only wants actual vs code users to use it.

[desert dune]

Not really. You can open it in the vs code terminal to the side of the code, and links in the terminal rerouting to the actual code in the code editor.

[desert dune]

Depends on how you want to use it. It doesn't assume full vibe coding.

You can also usually see a diff of the changes right in the cli(or whatever other agentic tool you're using). But I'd advice having git setup, at least locally, as it makes it a lot easier to track the changes.

[desert dune]

Yeah, you can do that. But I personally didn't like jumping between the ides.

[floral bobcat]

I'm currently setting up a local repo

Never done this before

I'm kinda getting too used to using plastic. I wonder if plastic can do local repo too

[floral bobcat]

Yes it can. Aaah so convenient

[floral bobcat]

https://learn.microsoft.com/en-us/visualstudio/ide/reference/command-prompt-powershell?view=visualstudio
I guess i set it up like so (i use VS)

Why are all the install tutorial uses VS code..

Somewhere i see that they recommend(?) also installing gemini CLI companion
Searching that and it's all for VS code
Is there no such equivalent in VS?

There is this extension for VS
https://marketplace.visualstudio.com/items?itemName=JocysCom.JocysComVsAiCompanion

[desert dune]

Generally vs is more rigid with extensions. But if it's a cli tool, I guess you can use it from vs without problem.

[desert dune]

This doesn't sound like anything useful.

I heard vs 2026 has a lot better integration with tools like github copilot and such, so maybe try that but yeah, it's probably gonna get any close to gemini cli on the free tiers.

[floral bobcat]

Im using VS integrated powershell terminal with gemini

Literally just starting out so haven't done anything yet tho

Without that VS AI companion extension*

[floral bobcat]

But why wouldn't it?
Altho yes, whatever this thing claims to be able to do, pretty much that's what github copilot/Gemini cli already does
I dont really get what he meant by better seamless integrated workflow than github copilot tho

[desert dune]

Well, for starters it hasn't been updated for over an year. Even VS/microsoft default tools probably got more features by now.
And overall, stuff like this embedded picture don't really make it seem like a solid extension.

[floral bobcat]

It does say it doesn't aim to replace github copilot, but to work along with it

[floral bobcat]

Yea ok for this particular extension i also have some doubts
But as someone who haven't tried proper agentic workflow, don't they say that they work with multiple agents handling different roles/scope, which is what this extension said can assist with

I do plan to also try the unity mcp to handle the editor side

Like im currently making custom postprocessing. If i need to quickly set up the scene to make an ideal setup to test for DoF postprocessing, an editor side Ai can help with that right?

And then it's a matter of getting the IDE AI and editor AI to communicate with each other

[desert dune]

You simply don't really need any extra tools beyond your main agentic coding tool(codex, opencode, gemini cli, claude code). They basically provide all the futures you might ever need already. Things changed a lot in the last year, so what that extension provides is already largly obsolete or is integrated in the tools I mentioned.

[desert dune]

With unity mcp, any of the mentioned can do it.

[floral bobcat]

I'm now just on gemini-cli on VS terminal
From /stats

Session Stats
│  Interaction Summary
│  Session ID:                 
│  Auth Method:                Signed in with Google 
│  Tier:                       Gemini Code Assist for individuals
│  Tool Calls:                 21 ( ✓ 20 x 1 )
│  Success Rate:               95.2%
│  User Agreement:             100.0% (21 reviewed)
│  Code Changes:               +316 -269
│  Performance
│  Wall Time:                  23h 37m 35s
│  Agent Active:               16m 9s
│    » API Time:               16m 6s (99.7%)
│    » Tool Time:              2.8s (0.3%)                           │
│  ────────────────────────────────────────────────────────────────────────────────     

│  ────────────────────────────────────────────────────────────────────────────────                                                                                                        
│  gemini-2.5-flash-lite     11    ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬    1%  1:04 AM (19m)
│  gemini-3-flash-preview    34    ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬   20%  1:11 AM (25m)
│  gemini-2.5-flash           -    ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬   20%  1:11 AM (25m)
│  gemini-2.5-pro             -    ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬    0%  12:46 AM (24h)
│  gemini-3-pro-preview       -    ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬    0%  12:46 AM (24h)
│  gemini-3.1-flash-lite-…    -    ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬    1%  1:04 AM (19m)

Sry guys i'm so noob. Just wondering about a few things

From a single prompt, i get response that can have multiple of these [Thought: true]
Does this mean each of these is a call/request/inference thing?

[chrome apex]

a single prompt in today's agentic-tools will cause a lot of sub-tasks (thinking), tool calls and a variety of sub-agents or alternative model uses that match the task.

a single prompt can cost you from 0 to $3 (equivalent) depending on what you tell it to do (thats a typical cost-safe default-config most agent tools have). if you have non-standard configs that can be much more and consist of multiple steps in which the agent will re-prompt itself until the original task is completed.

usually > 50% of your quota will be consumed by tool-responses and parsing of those.

[floral bobcat]

What do i do about this? Stuck for 21 minutes and still going
Also, i notice gemini barely ever asks me to confirm/validate things in the editor side. Back when i was on chat gpt, it sometimes does tell me to "now, when u try param x to value y, u should see abc" etc
I have the same context prompt(?) thing in its gemini.md

And this text is still changing around. Now it's
Delete from the cursor to the start of the line with Ctrl+U…

[desert dune]

What do you have in gemini md?

And yeah, gemini tends to over think or get stop in thinking loops. Might want to lock the model to 3 flash preview, as it might be using one of the older models that were pretty bad at it.

[floral bobcat]

I barely read all the output when they make code changes. Some thing's not working and i just kept saying "no, still not working"

Now i'm gonna read the last thing it says actually

My gemini.md

Unity 6.3, URP 17.3, Forward+ rendering.  Targeting webgl
All the API used in these codes are already correct, so don't try to use some other approach. Stick with DrawProcedural in the render feature, GetFullScreenTriangleVertexPosition in the shader, etc 

If need more info, ask more questions about the context

After getting my approval about ur understanding of the codebase, then u may explain the bug, correct code, explanation of correction, etc

I'll post scripts, as part of a bigger codebase
Will also post a question or issue
Do NOT respond with a solution before u have enough info. U MUST ask for more relevant info/code from this codebase, when the class is called etc

The bottom part is the general one. But it's more for chat gpt where i post parts of the codebase instead of it already having access to the codebase

[floral bobcat]

I might try claude now

My vibe coding friends are not using gemini and using claude insted

Still not sure what's the difference between uhh the LLM and the uhh hmm the.. ??

Use OpenRouter & free models to run Claude Code for free forever without hitting rate limits or paying $20/month.

[floral bobcat]

Like i'm reading this. So what's OpenRouter and what's Claude Code?

[chrome apex]

Each provider has a API by which you use its models. OpenRouter is a broker API that aggregates ‘all’ these vendor specific ones into a single one. Claude Code is an agentic coding assistant from anthropic.

Worth noting based on what you said above, relying only on free models gives you a decidedly different experience than using the paid premium models (depending on how ambitious your task is)

[floral bobcat]

https://prince-arora-aws.medium.com/how-i-set-up-claude-code-for-free-no-subscription-no-credit-card-and-what-i-learned-along-the-2cba880682a2
I wonder if this is what i want

I just have ugh annoying experience with VS code so i just try to make it work with VS

[floral bobcat]

I mean, i dont think what i'm doing is super complicated. I try to use AI just for specific features, and not so much about architecture or...
Actually what is considered ambitious and what's considered not?

I don't think i'm doing agentic multi agent vibe coding yet. I'm happy with just 1 agent for now

[desert dune]

This is quite outdated, but I'd probably agree with correction: Google models(they mention gemini 2.5 flash, but 3 flash is the right to use today) is probably the best you can get for free. While there are providers of free Chinese models in practice they lag behind frontier models.

[floral bobcat]

Damn that's 1 week old and it's already outdated

[desert dune]

Is it?

It's talking about models that are maybe 1 year old and are being replaced already.

Perhaps it's not the article but the author that is outdated. 😅

[desert dune]

You can add an instruction to your gemini md: at the end of every task, explain how to test it manually in the editor or something like that.

[floral bobcat]

True can do that. I mean the editor side test is not that hard to do. I just play the sliders etc
I'm not actually sure what i expect him to do/verify actually.. and if i explain it'll be too specific to the problem

Anyways i'll read its output first to see whats up

[floral bobcat]

The article is more focused on how to use claude for free

[desert dune]

I guess so.
But then if I were to choose, I'd use opencode instead.

Or rather I already use. Alongside codex.

[floral bobcat]

Maybe i'm just eager to try claude, but i still can only do free one for now

Btw still reading gemini cli's monologue. I see like "Let's ask the "investigator" to check the logs or something? No, I must infer." So i guess it's already doing a multiple agent kinda thing?

Sry i'm so new to this, just wanna understand what's it doing

[desert dune]

What you're gonna try is claude code. Not Claude. There's really only "code" in there. No Claude at all.

[desert dune]

possibly.

You only really need to read the last overview message. No need to read it's thought process. Unless you really want.

[floral bobcat]

And, some also mentioned maybe i should just start a new session. This last particular monologue is so long. I've only read a few lines at the end and now starting reading from the top too. And among others that it's trying to figure out some texture format n stuff, i guess what i'm missing (that my chatgpt does) is it would ask some confirmation/verification like "can u make sure the format of the texture in frame debugger for pass A and B is abc?" and those kinda questions

The monologue of this gemini cli is doing question answers with itself and the "investigator" and maybe some more
It (i assume) doesn't have access to what's set up in the editor (i'm not using unity mcp yet so maybe that'll help a lot in this case), so i reckon relying on me as the investigator is better than it doing Q&A entirely with itself

[floral bobcat]

Like why would it insist to "No, i must infer" ?

  I'll add a "debug" output to the shader. Like return float4(1, 0, 0, 1); (Red).
  But I can't ask the user to do that efficiently.```
Wtf? He really doesn't trust me here?

When i was using chatgpt, it would do all these kinda test and i'd happily cooperate. I still can't say which one's better overall but i think these self monologue/investigation is very prone to going so far to conclude something that should be non issue

[desert dune]

Add an explicit instruction: "For information you can't infer from the codebase, ask the user for verification in the editor/frame debugger/profiler/etc..."

[desert dune]

I think one issue is that you have a lot of "I" in that prompt. From the perspective of the model, this is just an instruction written in a book. It doesn't necessarily associate it with the user. Which is why you need to write it in a more formal and structured way. Instead of "my approval", "user approval". "I'll post scripts, as part of a bigger codebase", "Will also post a question or issue" - this is not needed at all, it's just noise for the model. It might even confuse the model. It can probably get the bigger picture better than you. Instead of providing it with code, provide it with intention.
"Do NOT respond with a solution before u have enough info" - you don't need that with agents, as they're in research mode by default, and this instruction probably causing it to NOT ask you follow up questions or check stuff in the editor.

Also, the grammar and punctuation are terrible. Arguably, this can affect how well the model follows these instructions.

[chrome apex]

the effectiveness of any model/agent lives and dies by the context and instructions you give it. if any of those are contradictory or too specific or in any way relying on judgement that isn't embedded in its training data you will get nonsense during chain-of-thought/reflexion/RAG/Reasoning. if you're operating in low-cost-models you'll get a lot of nonsense because the models can't really hold much context, therefore they constantly forget what you just told them and they compress the wrong bits. Its worth checking out what you can get from a frontier model with 'high' reasoning just to see a benchmark of what kind of understanding is possible before you develop a false intuition from working with the low-end. In practical terms, low-end models/agents are generally a waste of time for anything beyond translation/formatting/boilerplate tasks, and event there they are barely worth it. the first truly effective model for Unity was Opus 4.5 and GPT 5.3, those actually managed to get to the end of a serious bit of code that eventually looked like it was written by a sane human mind. Models before that did fail due to drift/churn before the code reached meaningful problem-solving power. they worked for throwaway tools and small features, but no big things that need a long dialogue of human-in-the-loop adjustment and evolution.

[desert dune]

I must disagree with that. Earlier models could already do meaningful work. Gemini 3 flash was specifically good(better than pro, due to pro getting stuck in reasoning loops) for its cost. I don't know about anthropic models, but gpts generally were able to do stuff end to end properly since gpt 5 codex(or what was the earliest 5 codex?). At least from my experience. But it depends heavily on what kind of task you throw at it and how good your specs/requirements are. It still does. And probably will forever. Because the main issue is not in AI. It's getting the ideas out of your head and conveying them to AI. I like to think that the bottleneck in AI workflows is the human actually. If you tell it something that doesn't make sense, it would produce something that doesn't make sense. Though models becoming a bit better at "telling you that you're an idiot". Hopefully, that area keeps on improving.

[chrome apex]

obviously this is on a spectrum and depends on the specific task, my specific experience pre opus 4.5/codex 5.3 was that the point of failure of the agent-aided process was so late in the development of a module that the sunk cost was really painful (weeks). the same task was eventually solved with opus 4.5 in two days, codex 5.2 before that came close, plain 5.0 caused the most wasted time. the spec/context didn't change, so i'm pretty confident this assertion is correct (in the context of unity at what i consider "significant"). The task was a from-scratch rewrite of an existing v1 of the module that originally took approximately 3 months of human time. the problem-space/spec was quite well understood.

the problem pre 5.3 was that the agent would run into such significant churn/drift issues that they could not even detect them fully anymore. since their mapping/debugging abilities still are very limited the 'breakthrough' in 4.5/5.3 (and even more so 4.6) was that they would basically solve the problem in one pass without any major opportunity for drift, each iteration became essentially a full reconsideration of every aspect, that was not possible at this scale before. Pre 5.3 to build something of significance involved breaking down the work into many incremental phases which was very slow eventually broke down for said reasons.

i find the 1-3 month of human dev time to be roughly at the point where a problem gets large enough that a person cannot fully trace the code anymore, meaning they begin to speculate and trial&error certain changes. so solutions at that point need to deal with the limitations of what one can still reason about. Any changes to such a feature/module take a lot of manual time and this is then where the agent help becomes impressive, due to its diligence and the ability to not simply refactor a few names and signatures but to reshape a whole architecture in half a day. This would have been impossible to do before for economic reasons.

[floral bobcat]

I'm still waiting for the agent to call me an idiot. There's a specific behavior i want that, i think does sound like "i want a frozen water but make it hot" situation and the agent doesn't seem to be able to pause and question my plan further before continuing to attempt some stuff
I did add the instruction in the gemini.md

[desert dune]

Maybe try planning mode then

[floral bobcat]

Ah true

[mortal eagle]

Thats the "walking to the car wash" problem. It's kind of a fundamental issue with LLMs. They are designed to give you the most likely result, if that happens to be .001% confidence then so be it.

[floral bobcat]

Btw guys. When an agent reads codebase, can it identify method calls, refernces etc? Or they need to read the whole everything to track all the connections?
Would it be better if i also manually provide the references for the agent to check further for the relevant scripts?

[desert dune]

You could provide more context, but agents are generally very efficient at navigating the codebase. They would usually search by keywords and only look at relevant context around. While search by references could theoretically be useful, I think it wouldn't work that great since models are note trained to use it efficiently.

[chrome apex]

they don't have something internally akin to "static code analysis", they may however call static analysis and refactoring tools that can make symbolic transformations, other than that, they do what dlich said, read the actual code and guess what it means. From those guesses, they read more code and guess more. They will generally use static analysis to catch errors in their guessing and correct them, so they will eventually arrive at correct code they present to you. This is also why a tool supported agent in an IDE or other environment with tool support is much stronger than one that doesn't have those analysis/testing tools.

[floral bobcat]

I will use glob to find files containing DoFBokehSquare and Character to avoid potential issues with grep_search and more accurately locate the relevant classes.
So that's cool
What's glob tho?

Anyways it does manage to do this so all's good