David Szabo-Stuban @screenlessdad

all those super convoluted Openclaw and Hermes setups are just a bunch of Markdown files in a folder structure. it's the equivalent of throwing more money at the roulette table hoping you'll get it right this time. and the agents will still fuck it up. if you want to make your agent smarter, improve the harness not the prompts.

designing a system for zero ux is harder than i thought. a few issues i usually encounter that kill the “fire and forget” approach I’m building Alfred for: tasks that are supposed to fire, don’t. this is usually some system health issue. an OOM here, a silent crash there. been slowly getting rid of it one by one tasks fire but the agent sandbags them. this is heavily dependend on the model we use. ie: Alfred sends a daily brief with tasks. I respond with comments on those tasks. then Alfred claims tasks were updated but in reality they weren’t. the model does the heavy lifting here. Opus 4.7 is best but can’t use via Claude Max. GPT 5.5 is close. Kimi K2.6 and Minimax M2.7 are terrible even though they are pretty close on benchmarks agent executes with insufficient context: this is the hardest to solve. we need the agent to always have a “minimum viable context” composed for them right before a task is fired, ideally done programmatically, at runtime. too much context = better results but expensive. too little context = bad or tone deaf execution the big solution that seems to work well is the Matter <> Task <> Decision trifecta. In Alfred’s vault there are projects, organizations, etc… but for the user, there are Matters which are ongoing concerns. think active projects without clear deliverables that evolve on the fly. Tasks are surfaced from data and then matched with Matters. Tasks also evolve on their own via a mutation log. if new data arrives, the task gets updated, if a task gets updated the corresponding matter gets updated too. Alfred can close tasks but can’t close matters on its own. the third missing piece was Decisions: whenever Alfred needs human input, it will route it to you via a Decision. your decision decides what Alfred does next and the decision as well as the outcome all gets logged in Task / Matter mutation. this system allows Alfred to just pull the right context from the task and the matter files only without overfeeding itself with context. seems to work, now stress test

Here's how Alfred is onboarding itself with (near) zero config. youtube.com/watch?v=_Y2K5-…

just gave my agent a card @agentcardai agentcard.ai/?a=screenlessd…

AgentCard @agentcardai

3 months ago

Introducing AgentCard. Your agent can now buy anything: • pay for inference & APIs • order DoorDash, Amazon, Ubers • run marketing • trade Polymarket 24/7 Open to all, not just businesses 🔥 Instant. Private. Reusable. Live today.

271 280 2K 1.1M 1K

Alfred now onboards itself. I started a Design Partner program and started onboarding a few people where I build and configure Alfred for them myself, personally. One of the first bottlenecks was onboarding. How do you get an agent to know enough about you to be useful? Alfred's onboarding process now does it all automatically. Click a button and within 30 minutes you get a fully personalized agent that knows you deeply without you having to lift a finger. It's a bit crazy tbh.

a few weeks ago I realized a major issue with how I build Alfred: it's a hot mess of duct tape, wd-40 and bad engineering practices. not because i don't know how to build it properly but because if i build it only for myself then i'll always just opt in for quick hacks and fix later. so as a forcing function i opened a waitlist for people who would want their own Alfred's. if i do it for others too it MUST work properly. that is Alfred Black (alfred.black) and for now I will work closely with a few design partners to *really* make it work for them. i was reluctant to turn this into an offer because i'd rather spend time adding new features than answering support tickets, doing marketing and all the other stuff that comes with running a business. but i realized that i will never be as diligent for myself as for a client so maybe taking money for this service will make my git worktrees nicer and the architecture cleaner :D to my surprise, I sent out a few emails on friday and my calendar filled over the weekend. anyway more on this later.

local llms are the future.

Qwen @Alibaba_Qwen

3 months ago

🚀 Introducing the Qwen 3.5 Small Model Series Qwen3.5-0.8B · Qwen3.5-2B · Qwen3.5-4B · Qwen3.5-9B ✨ More intelligence, less compute. These small models are built on the same Qwen3.5 foundation — native multimodal, improved architecture, scaled RL: • 0.8B / 2B → tiny, fast,

908 3K 21K 9.0M 14K

Alfred is growing up. Over the last week I started working on a cloud deployment stack that would make Alfred secure, scalable, portable and intuitive. I'm doing finishing touches now and below is one part of the new dashboard. 1. Alfred's Brain is now visualized as a real-time graph following the ontology. 2. Whenever the Curator, Janitor, Distiller, or Surveyor is working, the impacted nodes glow up and start pulsating. 3. You can run QMD-based search on your vault content and the relevant nodes light up like a Christmas tree ...honestly, this is just a gimmick, but I grew up on Imperium Galactica 2 so I kinda wanted to build this...well....Galaxy of Memories. Anyway, over the next few days I'll share some more info about the upcoming *Alfred as a Service* project:)

the Screenless Dad passed 3000 subscribers in 37 days. apparently lots of people want to be more present without going amish. the whole concept of "treating presence" as an engineering problem started in 2024 when i realized that i could probably offload most of the mental junk i get from device use to an agent. i realized i didn't want an agent or a chatbot. i wanted something that runs in the background, attends to the small stuff so i can be present for the big ones. to delegate attention so my mind frees up. aristocrats had butlers for such matters for centuries. butlers maintaining entire households. it took me a LONG time plus a bunch of radical progress in the industry to make it happen, but Alfred is finally becoming something truly useful. so i shared the how to with my subscribers this week (it's on the Screenless Dad) and I got ~130 stars on Github so far. i honestly don't know if that's a few or a lot. that's the first time I shipped something completely by myself and i'm just happy as hell about it.

@horstenegger @Hesamation @kepano @obsdmd world's Alfred's assemble!

@jameesy i'm loving obsidian so much. i could never be disciplined enough to maintain a good vault so i built my agentic butler's entire, Alfred which captures, groups, links notes together automatically github.com/ssdavidai/alfr…

Alfred is the difference between “OpenClaw is a techie project” and”our family has a private agent”. I’ve been working on it since 2024 and today I published it. Fully open source, free to anyone. Full tutorial available on the Screenless Dad. I built it for myself and I will keep sharing what I build. Alfred is becoming a no interface agent. I don’t have to prompt I don’t have to use a device I just…talk. Not a phone call or a voice mode, just braindump into thin air. Then the todo list appears, the reminder is set, the florist is called and flowers are ordered. Alfred is doing some of that and I will keep working on it until I can go fully screenless because Alfred willcbe my only interface to the web. It kinda went viral and lots of people reached out asking to deploy Alfred for them. So I’m launching Alfred as a Service Installation, full maintenance, the integration of Alfred into YOUR life. Not an agent, a butler. Not sure if this will become anything at all but I just opened the waitlist and you can apply if you want Alfred for yourself. → stan.store/lumberjack

@koylanai OR, what if traverse the skill graph metadata with say python and compose ephemeral skills at runtime so the agent only ever gets one skill

@karpathy not to mention that progressive disclosure of cli tools makes it superior to the context bloat of mcp tools

@asaio87 yeah but what about an unserious business?

@karpathy @korbencopy same. i'd need 10DGX Sparks to make it remotely close to useful with local compute

Alfred now turns OpenClaw into a butler you can truly forget about. Alfred's vault now lives works on Zo, macOS or any Linux VPS. the full automated pipeline works even if you don't have Obsidian installed. it also have an execution engine (I use that instead of cron) that is based on Temporal (you need to host temporal locally though) doing some bugfixing today and will come out with a massive deep dive walkthrough tomorrow. github.com/ssdavidai/alfr… install with: pip install alfred-vault