Seriously, it’s feed and dashboard tools — underrated and not talked about yet. It’ll build artifacts and personal sites for you on the fly, it’s fast, but the web research and actual programming ability it has is limited. It’s best signals are your emails and social networks and messages if you’re into giving it that.
It's hard to wrap your mind around at first, because it's honest to God, supposed to be how most people implement agent use. Just fire up a routine that reads email signals and addresses items with tools you give it to address those items day by day as they pop up, and you'll have an agent more capable than anything else you'll find on the market by the end of a week.
Yeah, I’ve found when it tries to jump in to help a few well documented repos and new routines/builds I was working, it just messes stuff up and won’t follow all of my explicit instructions given a detailed prompt.
Muse spark 1.3 is supposed to be good at long running agentic work, but I find it terrible. Well muse is anyway.
I could see my giving it to a few cheap and easy routines that currently struggle on local model strength and smaller context windows, but shoot, I’m worried it’ll have lackluster solutions or lose some of the context it’s working in, that requires frontier intelligence to follow up behind it and clean up after it.
I’ve tried running the routine on Luna max for example. Not enough to keep stuff from falling through the cracks, or the agent stopping mid task if it encounters any issue. It requires a Sol/Astra/Opus 5.5 session to fix it constantly.
Would muse spark 1.3 cut it? I’m sure it’s better than medium weight… but I’m not sure it’s much better than Luna max unless someone can convince me otherwise.
@andr3barroso@raywongy No but that’s the thing, people average 6th grade reading levels, even the college graduates.
LLMs - MASS TEXT GENERATORS are literally THE WORST tool to give human beings in the year 2026. 😂
Not a problem for those of us that can and do read, you’re right about that.
@senb0n22a@B_doong2daddy Yeah api billing would cost hundreds of dollars a day on the projects and routines I work on opus 5.5.
5.5 is much better on a Max 20x plan but cursor is pretty good about usage!
Even worse, he gave it permission to scrape and buy for him… and apparently forgot or didn’t even know what he’d done.
This is the problem with LLMs. People don’t read. Literally do not read. They give all this execution space to the ai, then don’t even pay attention to half of what it’s doing, then they’re surprised when it’s doing exactly what it said it would 60 messages ago. 🤷♂️
He’d inadvertently setup auto replies and a Facebook marketplace search and buy routine with the agent.
That’s the thing about LLMs. They spit out a lot of words. Sometimes they do things in those paragraphs of words, and you can’t just “yes yes yes continue no mistakes” until stuff like this happens. 🤷♂️
It’s less about doing inside muse, than it is accessibility for most people.
For example, not everyone wants to pop open multiple ai apps or tools, so they use muse as a delegator or chief agent.
I used to use it to brainstorm projects or tasks, plan the prompt and write it, then spawn and manage projects in CC or Codex, then I’d open the app and manage from my phone anywhere I am, or let Codex and CC agents and workflows I already have in place do the heavy lifting.
Today, I have grokbots that manage that more effectively, while Muse is a more person agent tool. It shops for me or puts in orders, I don’t navigate much on my phone or computer anymore for the mindless scrolling and shopping and endless UIs and doomscrolling.
Ai helped me stop those bad habits, cleaned up my inboxes to create more effective signals and digests. I’m thinking about giving Muse more honestly and paying even just $15 for it a month, it’s nothing compared to what I spend on eveything else!
@kimindiehacker@viticci They have their own terminals and folders or online repos they can hook up to, and their own terminals to install CLI for basically anything.
You can have it SSH to a local computer you have and access projects or files for you or do everything self hosted. It’s pretty cool.
You raise a point I’ve been thinking about for months lately.
Think about it — we have 100s of accounts saved in password managers, apps on our phones, websites we sign into at work, at home, etc…
We navigate ENDLESS UIs. Endless hours on screens with unique menus, places to put things….
Dude I’m tired. Tired of endless interfaces. Tired of pointlessly learning a new CRM or interface or product site just to abandon it in a couple years and start from scratch somewhere else.
I. Am. Tired.
I want to build MY APPS AND TOOLS FOREVER and have agents that build and maintain them. That’s it.
Business-wise, I don’t want to sit here with my phone out, awkwardly talking to the receptionist at the dentist figuring out if I have an appointment 6 months from now on a random day in March, while she awkwardly looks around an ABSOLUTELY PACKED SCREEN with appointments all over a calendar.
Instead, I want to get my phone out, put it next to a little electronic pad connected to the computer, my agent speaks to the business agent, I talk to the receptionist about my day, and then my agent confirms with me 1 of 2 preferred options (it knows me duh) in March and the business agent knows the open calendar rules.
When my agent picks a day and I say wait a minute… let’s change that to the other option, me and the receptionist should be laughing and smiling and talking already as this happens in real time. 5-10 seconds or less.
I’m building an actual agent that replaces me at my computer. It’s close. It’s just simple routines plugged into AI. THAT is ai, I think, to the world, once they can appreciate it a little more. Less human to computer activity. More human to human activity.
Hopefully it makes us more social for us all to have agents that interact and engage with the digital “necessities” of our lives. One can hope!
@nat_eur@AsherandEmber@thsottiaux@anthdm People out here talking about gpt6 sucks and they’re out here admitting they’re using it as THERAPISTS AND CHATBOTS 😭😭😭😭
they're making big progress fast. to think, they basically had NOTHING just a few months ago. Grok Build, the Cursor buy, grokbot, the engineering team, they're all great.
Grokbot is the brand identity for what Cursor was building all along. it's why the plans are so funky.
It would be silly to think that the next step isn't a full-featured Grok app release that fully incorporates the cursor agents and grokbot together. It's taking time.
But it's all coming together. Patience. October should be HUGE for 4.8, app updates and more!
Those with human training data and inputs from social media, online searches, WWW, etc., are fast and cheap and smart.
Those focused explicitly on large hardware investments and scaling, are frontier intelligence.
It’s a small but noticeable difference in terms of results.
But look at how fast meta is moving now? They train on the intelligence of frontier intelligence and are still faster, cheaper, etc…
Google won’t be long behind. Only a matter of time.
Ai tools are cool and all. But unless you’re an ai agency selling ai tools, that doesn’t mean they have all the TOOLS you do.
Use AI to do business better. Use it to make cool unique tools, skills, applications that simplify workflows.
You both have AI tools the same way contractors and tradesmen all have ladders and trucks and hammers and saws.
Not everyone has simple systems that work.
Start there.
Everybody complains about usage but I’ll sit here spinning out a few hyper targeted projects on Opus 5.5 Ultracode or Max and run it for 2-3 days before hitting limits.
Plenty usage for me to delegate it through grokbots or hop between it and Astra 6 for a couple days. They all work really well together honestly, it’s hard to not go back and forth between the different high level models and get different perspectives on everything or have different tests and edge cases worked through.
Yeah, the problem is, LLMs literally are designed to just continue to spit out words until it comes to your desired outputs/conclusions. It only gives you what you ask of it, with some added engineering and tech that may or may not be necessary and that is the trap people keep falling into. allowing the AI “to just do” without specifying what it’s actually supposed to be OUTPUTTING on a turn by turn basis, is like rolling the dice everytime with an employee, by the time you get something that vaguely resembles what you asked for, but didn’t specify in so far as details along the way.
Do you want to give someone explicit instructions?
Or vaguely vibe your way to desired results?
Victim blaming is a funny way to put it, because hilariously enough, we all go through it!! It’s hard to get it setup right man, I’m struggling through it too!