Also, I thought Jules was the "coding agent" they are working on. Now this is taking it over or is this like another case of Google self-competing?
Someone needs to take charge at this company with a strong vision, because they are all over the place and spreading themselves thin, which in turn spreads thin the customer/brand equity.
At this point, as someone who: - Has been writing Android code for about 13 years now
- Has collaborated with Google on stuff
- Lead Google developer communities and conferences
- Knows many, many GDE's and has discussions with them often
- Uses Gemini API for their product
I'm so damn confused. How is a normal customer expected to understand then?
- They have 2 SDK's for communicating to their Gemini API.
- The documentation is spread and thrown all over the place.
- Half the time I'm trying to do something I have to dig through their code to find how to.
- The features I really want are rate limited or available only to private testers.
- They have 3 coding agents now.
- Even thought they have access to my Google Account and my phone, their Gemini app is useless.
- I tried to do a basic thing (add a service account) in Google Cloud recently, which wasn't allowed due to default rules that are deprecated and are so confusing to change due to their confusing UX.
The only usable thing is the AI studio, which is a great tool for experimenting with diff models and improved the DX of getting a Gemini API key by a mile.
I'd say congrats on the release, but honestly this is such a mid low hanging fruit of a product.
The chaos you describe is actually a significant positive in research environments. It's not spreading oneself thin, it is diversifying and decorrelating ones' efforts. You can't centrally plan all innovation.
But for the interface between the customer and the research output, which is a software and product problem, that definitely needs a different approach.
Personally, I would rather Google did this sort of experimentation even if it is more confusing.
Or I could be wrong about this. But following NotebookLLM, it seemed like the team developing it had a lot of autonomy.
And if it isn't shut down, it is left in that terrible half-documented state, with confusing integrations and terribly integrated into the rest of the product.
Considering I'm confused both as a customer, user and a shareholder, I'd say the tactic isn't working.
Remember that Google operates at huge scale, so even something any other company would consider wildly successful (e.g. Reader) is a waste of resources for them. That means that if you're ramping up your product over the course of a year you're wasting time and money. Go big or go home.
They don't get a sweet bonus and promotion for helping another team improve a product, so why collaborate? Just create your own chat app according to your own team's vision/goals/available technology and release it and hope it gains more traction than the other teams' existing options.
This is the funniest thing to me. When you open the app, Gemini says:
"Hello, Vasco"
In the welcome screen. I then ask this amazing intelligence this question:
"What's my name?"
"I do not know your name. I am an AI and I don't have access to your personal information."
I know why it happens, but it's so funny.
They probably do, along with "pretend to not have any personal information about the user".
Gemini CLI works synchronously with the user (unless you YOLO) and in your own directory on your own machine on your own checkout.
Two different modalities.
It may be the same coding agent behind the GHA. I question the implicit declaration behind OPs critique: that all 160,000+ Google folk should offer a single coding agent to their billions of users (or whatever the TAM is for coding agents). This is akin to criticizing Google Cloud for having VM, Kubernetes clusters and AppEngine; superficially, these products solve the same problem.
FWIW - this Github Actions integration is close to my ideal AI agents workflow[0]. I don't want to metaphorically look over my agents shoulder as it works in a specialized, vendor-locked IDE. I want agents to work asynchronously, taking however long they need, and tackling multiple tasks, with PRs/CLs as the unit of work. Current models may not be up to the task of single-shotting this, but the task is parallelizable across multiple agent-instances & the best solution selected (climate change be damned). I suspect Github alone may not provide adequate context as it may be missing previous tickets and design documents & the back-and-forth on requirements, but it's a start, and I'm glad Google is exploring this path for agents.
0. I believe in this workflow so much I created a proof-of-concept project that reads tickets from Vikunja and creates PRs using Aider some weeks back.
[1]: https://github.com/google-gemini/gemini-cli/blob/main/docs/c...
But the the "sunsetting" of projects good or bad, random shotgun approaches to everything, super awesome islands of product that slowly get bled dry... it is a failure of management structure, not just management.
I don't know the guts of Google, but I imagine there are 500 VPs (or equivalent) each with their pet project, each trying to curry favor with the boss who sent an email blast to "go big on Gemini". It feels like many teams just dropped their old busted projects and moved on to the new hotness, to hell with the customers, consistency or revenue. The only metric now is "Gemini engagement".
This appears to just be a plugin where you do things on GitHub, that sends out notifications to gemini-cli running on cloud, then gemini-cli responds and sends notifications back.
Basically just saving you the hassle of cloning at a specific commit, calling gemini-cli manually, and then uploading the result manually.
Now if they could get Gemini (the LLM) to run on a GH Actions runner I'd be more excited.
Well, they do have a lot to spread. But yeah, intense amount of overlap.
If instead of Google search they made 3 products each called "Google Search", "Super Search" and "YaGoo!", they wouldn't be where they're at today.
Similarly I tried contacting some human support for billing issues but was denied because automated checks deemed me unworthy for consulting anything besides documentation pages which i didn't understood so i gave up and switched to another cloud provider.
- Gemini can't make me a calendar appointment between myself and another person for 30 minutes in the next week. Heck it can't make appointments yet. - it can't edit or collaborate on Google docs, just insert. I edit my docs in cline or Claude code as markdown and upload. - speaking of, I don't think they have a MCP for working with docs or sheets - Gemini is worse than a Google search at helping me with sheet formulas
There's all these unique places in googles ecosystem I feel they could/should be excelling at AI at. They're not.
Hell I noticed yesterday searching for my remarkable preorder from years ago that you can't exact string search Gmail anymore. Searching for remarkable was pulling up "amazing". They're just degrading all of their products to stupidity at a time when I and AI can use more powertools
Also, if you are a Google Workspace customer, you can connect your workspace to the Gemini web app. It can then search and manipulate your calendar and your drive. It will also summarize documents and a few other tasks. I have less use for this but it is far from "it _can’t_ make appointments".
gemini in calendar says it can't alter my calendar
In the meanwhile the rest of us are running around with a pole that has a chainsaw generator on it swinging it in every direction to solve problems. When I have these issues now I just have AI write tools to get around it. To search my email I just had ai crack my takeout files and look for the email. Was easier than figuring out the google product.
Heck, just yesterday my partner forgot the grocery list printout so I took a picture of it and asked gemini to convert it to a format where I could copy and paste it to a specific todo list app that was already shared with her. INSTEAD, Gemini dumped the list into Google Keep, albeit with terrible formatting. Didn't miss a single item, but did not recognize categories of item (produce vs frozen food, for example)
So my read on it is there's a lot of "rough around the edges" use cases which can be tidied with better prompting/context or just Gemini team prioritizing those things when they get around to it.
What they actually _need_ is a marketing team showing off useful applications of the releases more often. OpenAI is ALL OVER TIKTOK and people I meet under 30 on that platform don't even know gemini exists. In my experience, Gemini is better than chatGPT at everything you need to do, and it can do the things that the OpenAI marketing people are constantly showing off on TikTok.
A simple but tedious task that I wanted to do for a large document. Nope, Gemini says it can't do that. It offered to tell me HOW to do it though!
Is there something Gemini can actually do that's useful?
combined with separate plans to use the Gemini CLI, it’s an incredibly goofy situation
If something like this exist please educate me as this would make tons of products better.
I'm sure it's capable of doing those things, but they have it turned off because of the significant risks involved in automatically editing important documents like that.
All it needs though is a sandbox to execute the action in, and an approval flow for the user to review the changes the agent wants to make, or make revisions. Why does it have to be all or nothing? "Hey Google, schedule a meeting with x for next week when we are both available" "Google: OK, here's a preview of the calendar invite - do you want me to send it, or make changes, or cancel?"
instead I'm just pulling it out to other tools and using markdown import. I'm basically moving away from using google docs because of these issues.
I tried it in Slides to generate some slides from text, it was useless.
I tried having it produce a plot for me in Sheets and it was useless. (It does look up documentation but I could already do that.)
I haven't found any useful feature for it in Google Workplace; but I do get AI summaries of everything now, which I don't need and have to keep dismissing.
I haven't yet used it in Google Docs -- it's probably decent there for writing.
So i THINK this is what it IS:
A GitHub Action that can be included in GitHub workflow YAML files. It executes the Gemini CLI, passing in prompts, repo context, and event data (like issue text or PR diffs) to generate responses or perform actions. In other words: it's a wrapper that installs and runs the Gemini CLI inside GitHub Actions environments.
It can use GitHub's API (via tokens or apps) to read repo data (issues, PRs, code) and write back (e.g., add labels, comments, or code suggestions). It makes calls to standard HTTPS API endpoints for Gemini LLM" (via the CLI's backend interactions with Google's Gemini API)
I've not explored this use of CC yet, anyone actively using AI-assisted CLI in CI/CD? Not automated PR review but either to semantically pass / fail an MR or some other use of terminal-capable, multi-context mashup during CI/CD?
In this case, the "chat" happens as a comment on an issue or PR addressing @gemini-cli
Again, with the complicated subscription. Please just give us a monthly subscription for developers that I can pay whatever, and then use Gemini CLI, this github action, Gemini chat, Jules, etc. Just like Claude and their max subscription.
This would be a game changer for me.
Sorry, congrats on the release too. This looks cool!
The primary goals are promotions, bonuses and stock price.
If that's the case, last i checked they are doing pretty well on stock price.
I mean this has been enough to get my feet wet and have some fun with exploring agent-based development, no doubt, and I appreciate it, but I'm having a hard time crossing my experience with,
> generous free-of-charge quotas
as they say. It's not that generous if it stops working after 5 mins? (This morning literally a single sentence I typed into Crush resulted in some back and forth I guess it called the API a few times and it just rate limited-out. Fine, it was probably a lot of requests going on, but, but I literally gave it a single small job to do and it couldn't finish it.)
Meanwhile I seem to be able to use the Gemini web app endlessly and haven't hit any limits yet.
Glad I didn't add an API key. I've had friends who did and ended up with $xxx in charges because the models can't think or use tools properly.
> Delegate work with an "@ mini-cli" tag and the agent can complete a range of tasks, from writing bugs to fixing bugs
Why not follow Claude Code naming with this and just call it `gemini github action` or `run gemini`?
The low quality human-authored PR's that came in (due to the incentive we offered) combined with the fact that a draft PR could be made for pennies with AI made this concept dead in the water as far as I'm concerned.
The pain point of getting some attention and action on your opensource codebase is really no longer relevant, in fact the pain point seems to be moving to how to optimize the limited reviewer / maintainer bandwidth under the onslaught of proposed suggestions.
To this end I've been experimenting with a framework that builds PR's from the major agents and but with a focus on how to structure the tasks and review process that optimize the review => accept/revise cycle. If you're interested I've been writing up some case studies here: https://github.com/sutt/agro/blob/master/docs/case-studies/a...
The Gemini assistant will need to be several times better than the existing tools to even fractionally displace them.
Having a hard time imagining the GHA integration will be much different.
As Microsoft own GitHub and it's a competitor.
As an engineering manager with an AI budget, I'm always looking for better and cheaper tools.
I have a decade of engineering experience and consider myself fairly intelligent.
I still can't figure out what this is, who it's for, or how much it costs.
https://x.com/tomgara/status/1587640766696140800?lang=en
"It’s pretty simple: Google Meet (original) was previously Meet, which was the rebranded Hangouts Meet. Meet has been merged with Google Duo, which replaced Google Hangouts. Google Duo has been renamed Meet, and Meet has been temporarily named Google Meet (original), for clarity"
Isn't this a recipe for disaster, or is all the FUD around agents wrecking havoc getting to me? I love Claude Code, but it can be somewhat bonkers and is at least at arms length from doing any real damage to my code (assuming I'm following good dev practices, and don't let it loose on my wider filesystem).
We have tried out Gemini code review vs Copilot code review and Gemini is consistently offering better code review tips. It has officially caught multiple potential bugs, even a few that reviewers might have missed, so it's definitely been additive.
Observability looks way worse. Github Agent has a full UX built into the Github PR that lets you dig into the agent behavior. This requires you to egress text logs and make sense of it yourself.
Also curious about customization. Github just rolled out "agent writes its own instructions" https://github.blog/changelog/2025-08-06-copilot-coding-agen... which is super cool, how do I customize this one and teach it how to start and manage apps across my monorepo?
Yeah that's on you. Add a `copilot-instructions.md` file and configure the `copilot-setup-steps.yml` workflow to setup your environment. Both are supported more or less since Copilot Agent got released (though in "preview")
Most agents read `AGENTS.md`, I just symlink it to CLAUDE.md, and do the same for GEMINI.md
@artdigital is on the money here. Our quick tip for beginners is to use `copilot-instructions.md` (which we can now generate for you <3), but for more serious use, we'd strongly recommend adding `copilot-setup-steps.yml`.
That gets you a deterministic setup - and for many teams, it'll be easy, as you can just copy and paste from existing Actions workflows.
Github now appears to support defining setup tasks in a Github Action that runs prior to the agent, so that's the next avenue of research.
Regardless, the website agent will always be slower. My local is already running and fully ready to go so the ide agent can hit the ground running on any task. The website agent has to spin up a machine and install and build. It will take time.
so the agents you spawn don't use paid Gemini tokens?