AI agents

[Tim Macpherson]

 Following on from a recent post about AI, I  use basic chatgpt or Gemini. Should I be using AI agents such as Claude Code, OpenAI’s Codex, etc ?  My current AI usage is ok for specifics but awful at remembering things.

Yahoo Mail: Search, Organize, Conquer

[Jens]

What is your goal? AI agents running locally are usually a security nightmare (data deletion, secret exfiltration via prompt injection) so you have two choices:

1.) babysit them all the time and review, understand and confirm any command the agent wants to execute. Otherwise it might delete your drive accidentally and feels really sorry about it or it reads an untrusted website with some prompt injection sections which causes the model to do something totally different. As a middleground you can allow read/search commands by default but still review everything else.

2.) Run the agent autonomously in a secure environment to protect your local data. This can be a local VM, a local docker container (e.g. use devcontainers for development in general or just run the agent in a container) or an OS provided sandbox mechanism. Then you make only the folders accessible to the environment (and thus the AI agent) which are required. If your agent should use the internet for research then you must control network requests (e.g. allowed domains).

Depending on the code size and the task complexity an AI agent is relatively slow as it does a lot of research in the code base first before implementing stuff. So if you work on a laptop it might be annoying to wait for the agent to complete. So the last option is using agents in the cloud which also solves the security issue for your local data as the cloud solutions usually must clone your repository from somewhere.

-- J.

[Tim Macpherson]

Thanks, I heard the agents in cloud can be a security risk for proprietary code, but there are options to keep it private.  Thinking about it an agent would be overkill for my purposes, I just want something that doesn't have to be constantly corrected because it's forgotten my current stack. It can just about remember that I'm using Java GWT and D3 js, memory of a 4 year old

Yahoo Mail: Search, Organize, Conquer

--
You received this message because you are subscribed to the Google Groups "GWT Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to google-web-tool...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/google-web-toolkit/9585ff0f-015e-4344-931f-6610ae632644n%40googlegroups.com.

[Tim Macpherson]

Gemini just output that for a complex refactor where I'm stripping out pure java logic from a GWT client module incrementally in a correct order so the app isn't broken at any stage: AI  can't do that, agents or whatever. Maybe in 5 years, I have no idea

Yahoo Mail: Search, Organize, Conquer

To view this discussion visit https://groups.google.com/d/msgid/google-web-toolkit/546002399.7992174.1767199698829%40mail.yahoo.com.

[Jens]

How do you use Gemini / ChatGPT? Via some IDE plugin or just their website? 

For example if you use Github Copilot and use the official IDE plugin then you can store markdown files in your project which describe your stack, coding patterns, etc. The plugin will then read these files and add the content to the context of each chat session you start. That way the model always knows these information. These files are different for models / plugins, e.g. Gemini CLI searches for GEMINI.md files if I remember correctly. You have to read the docs of the tools you are using.

These language models are good at pattern recognition and pattern completion. They cannot read your mind nor have a mind. What matters is that the context passed to the model is good and that it can see clear instructions so it must not guess. Only then the output will be good. 

-- J.

[Tim Macpherson]

I asked chatgpt about GitHub copilot and Gemini CLI,  Re  complex refactor where I'm stripping out pure java logic from a GWT client module incrementally in a correct order so the app isn't broken at any stage:   Chatgpt also says: your instinct — “maybe AI can’t do that” — is correct. It uses an analogy:
AI is good at: rebinding books, rewriting titles, rearranging whole shelves
It is bad at: Knowing which single book is safe to slide now.
That’s human judgement grounded in system knowledge.

Yahoo Mail: Search, Organize, Conquer

https://groups.google.com/d/msgid/google-web-toolkit/a2cad606-9195-4a27-8e11-3ca5119f6748n%40googlegroups.com

.

[Craig Mitchell]

If you use IntelliJ, I highly recommend the Windsurf plugin.  It allows you to use any model you like.  ChatGPT, Gemini, Claude Sonnet, ...

And as Jens says, it reads your project, so it knows what you're doing, and it's scarily good at helping you out (refactoring, or doing anything).

Privacy and keeping intellectual property private probably goes out the window though.  If you need that, you'd need a sandboxed Copilot subscription, or something similar.

[Vegegoku]

You can also use codex in the terminal with a persistence context, you start your first session and make it try to learn and understand the project context, draft an md file from what it learns, review and refine, repeat and update the file - for example AI_CONTEXT.md- until you have a solid starting context, add an instruction in the file to update it with new learnings an findings from every subsequencce sessions, add a shell script what ever you like in your project to load codex with that file as context, from that point on you will have an incrementally built persisted context and you dont worry about the agent forgetting all the context you had before., this might also work with any agent as long as you can invoke it with a pre-defined context.

I found this useful for my case. 

[David Nouls]

Windsurf can guarantee the privacy. The company where I work is very strict on what we are allowed, but they allow us to use Windsurf. 

 

It is quite impressive at times, but it is not very consistent. One day it blows my mind, the next day (with exact same prompt) it delivers something stupid.

 

But it does help me a lot in trying to out new ideas as long as you are willing to put in extra work to polish it into something usable and maintainable.

To view this discussion visit https://groups.google.com/d/msgid/google-web-toolkit/b648b7ce-f176-4cf8-820a-7521dd45c4d8n%40googlegroups.com.

[Tim Macpherson]

Thanks all for advice, I'm looking at Gemini cli today with non-agentic commands - don't do anything without asking. It has a generous free tier unlike Codex. It should solve the memory issues with a single Gemini.md context, but already it's done  a 4 minute analysis ignoring the stated dependencies. Generally an improvement though, wondering if $200 pm Codex is worth it.

Yahoo Mail: Search, Organize, Conquer

To view this discussion visit https://groups.google.com/d/msgid/google-web-toolkit/adaca253-b15f-44c0-99fc-6a937f960e52%40Spark.

[Vegegoku]

When I mentioned codex I was talking about this https://developers.openai.com/codex/cli and from the pricing page it is available for 20$/month subscriptions.

[Tim Macpherson]

The $20 pm codex is not for full-time usage, but not sure if I need full-time, so far Gemini cli is better than eclipse at tracing references, haven't done much else with it. The UI is awkward compared to webpage AI,  I'd guess most Devs use both, webpage  chat with .md context file would be useful but I haven't heard of that

Yahoo Mail: Search, Organize, Conquer

https://groups.google.com/d/msgid/google-web-toolkit/2986385e-c948-4b8b-a4e1-b292d656698an%40googlegroups.com

.