currently what i do is setting up a group for a set of sessions/agents, once this is done, i'll name the group based on the task/activity which is being completed.

by doing this i can move between several tasks, see the current status of each session, and even use Blueprints to send a message to a certain group of sessions/agents

Hey, I am more than happy to answer questions. Please try Blueprints with Seshions, I used to have to spawn the same 4-agent setup every time manually. Now it's just one keypress.

This sounds as a great opportunity to use this as an RL env to train drones for defense purposes.

Thanks for your comment. I agree. If expanded, it could serve as a test-bed for experimentation and training purposes. Educational interest is also one of the possible goals. Feel free to try it out or send to interested people. Cheers!

How does this compare to using Nano Banano or Seedance 2.0?

"Great question. The key difference lies in the underlying alignment philosophy.

While models like Nano Banana or Seedance 2.0 (and certainly those from Google/OpenAI) have moved toward extremely conservative safety layers that often result in 'over-refusal'—even for benign creative prompts—Grok Imagine 2 is designed with a much leaner guardrail system.

It prioritizes creative agency and unfiltered expression (consistent with the 'truth-seeking' mission of xAI). In practice, this means:

Less 'Preachiness': It doesn't lecture you on why your prompt might be 'problematic' if it's within legal bounds.

Nuanced Realism: It’s more willing to render gritty, cinematic, or edgy aesthetics that others might flag as 'unsafe' due to strict corporate brand-safety guidelines.

Contextual Freedom: It understands satire and historical context better, rather than applying a blanket 'no-go' policy.

I definitely am going to give it a try, it's been a while looking for something like this. Had an experience with Cap, but when you use the camera, the recording breaks

One of the use cases i see for this tool is helping companies to understand the output coming from the llm blackbox and the process which the employee took to complete a certain task

Why do that? Just let them deal with it.

Except it doesn't capture the majority of uses of AI, in my experience. In my current practice, the the vast majority of AI use is autocompletions, or small inline prompts. ("Fix this error."; "Open an ALSA midi connection" (things that avoid a to trip into awful documentation); "if (one of the query parameters is "gear='ir') ..." (things that break flow by forcing a trip into excellent but overly verbose Javascript URL API documentation)). Only very occasionally will I prompt for a big chunk of code.

What coding agent are you using at your workplace?

Claude Sonnet 4.6. With some brief experimentation with Gemini 3 Pro and GPT 5.3 Codex (both of which seem excellent so far). Perversely, I'm more comfortable at the moment with Claude Sonnet 4.6, just because I have a ballpark sense of how far it can be pushed before it loses it, and haven't yet developed the corresponding bat-sense for Gemini 3 Pro and GPT 5.3 Codex.

I've used Ghostty in the past, but nowadays it's hard to use it since you have to take in count that running agents requires using tmux sessions which is something Ghostty does not support very well

I went through the article, and it makes sense to me that we're getting names as an output, but why doing so with names?

Names is just a random problem to demonstrate the model. It could be anything, I believe

What if you just use words

Probably because names kinda obfuscate the ridiculous impracticality of this exercise. This microgpt can produce a random sequence of letters and by chance it might look like a name. If the thing output, let's say "Kianna" you just think "wow, it IS a name" but is it though? (Idk if it's a real name, at least not in Spanish) Isn't a normal word, so the randomness of names helps to hide the fact that this gpt just outputs random shit that looks like names. If you just use words you will get mostly random shit that doesn't resemble any real words. Just my hypothesis. I can see the convenience of using names. The output look like real names but you can achieve the same result with old ai and very basic algorithms.

in other words, it works well because it doesnt make out weird names.