YB

@signulll I’m a huge fan of how @nayafia defines agency. Can you, given your current conditions, change your circumstances open.substack.com/pub/engineerin…

If you're using agents, make sure to be thinking about your sovereign memory. The built-in memory and user md files should primarily be mappings to your local memory setup. This can be obsidian or a simple folder or google drive doesn't matter. What's important is that you're adding continuous context to an endpoint that you think of as 'home base'. Your agents need a working surface (it's like giving your employee an office). Don't stress too much about organizing your memory setup, skills, cron jobs, etc. You don't need to figure out a perfect strategy for how your agent will work with your home base from day one. Keep it simple stupid. The key point is to start adding as much as you can too your sovereign storage ASAP! Why? Because context compounds. The more clippings, conversation threads, voice notes, etc. that you add now...the better you'll be able to use your system as the agents continue to get better and better! The most important thing is to capture lots of raw context into your durable home base first, then let better workflows emerge from daily use. I would even argue that it's better to be overkill when you start and then find a happy medium.

@travers00 @paraga @benthompson oh wow, funny timing. thanks for sharing!

@yb_effect @paraga might find this pod with @paraga and @benthompson this morning interesting stratechery.com/2026/an-interv…

Yes! For humans, ads was the most efficient business model due to the human cognitive overhead with payments. With agents, direct payments will work much better. This is why I’m so optimistic about high quality content thriving on the web as agents take over.

@0xLewis_gg Ah that’s fair I guess. But I’m still using just Claude / codex so I’m assuming it’s more enterprise facing agents?

@yb_effect it's all from agents, and half of the usage on openrouter is still those same frontier labs

realizing I want to better understand how people outside the Anthropic/OAI/DeepMind SV bubble are using and thinking about AI. my feed is too frontier lab congested (that's great but too closed off from reality)

@yasu0x1 any good examples you can point me to? like which model is good for which niche use case?

@yb_effect small finetunes doing the real heavy lifting that never gets clouted half the useful stuff is on 3b models nobody has heard of

Creating is becoming easier. Getting people's attention is getting harder.

@ThisRikka examples of niche tasks?

@yb_effect the real action is happening in shittier models getting productized for niche tasks no one at SV would bother with

reminds me of the classic crypto chain routing dilemma / DEX aggregators. but of course there's just a lot more variability in the tasks for general agents. if models do continue to commoditize AND people are okay with non-frontier models for the long tail of tasks, then I think the most interesting thing to think about is how the model routers (agent orchestrators or whatever you want to call them) get paid. my guess is a hybrid of how many tokens each intent cost as well as how successful the task was done (time, accuracy, efficiency, etc). this is great because now most people don't have to worry about subscriptions and can simply monitor the allowances we're giving. forget normal consumers, most people in tech have stopped trying to keep up with all the OS models out there. i myself have subscriptions to gpt, claude, and grok and already feel overwhelmed

@mignano yep, i think incentives are better aligned. and it can get more interesting if there's "agent insurance" or "agent taxes" as well that can be incorporated at more micropayment scales as well

@mignano your agents will either be big insurance or amazon customer service x.com/1ssve/status/2…

@yb_effect oh that's a great call. i like the insurance idea a lot...

@yb_effect yea i think the most obvious way is to charge a margin on the tokens but that feels like not a great place to be. i like the idea of rewarding performance somehow and attaching revenue to that.

they need to read more like recipes, less like classroom tutorials

how agent friendly are your developer docs?

@silver_pump think it's the classic "walled gardens vs open web" argument except for agents

@yb_effect wait so youre saying every agent stack eventually becomes the OS it runs on idk whether thats progress or just history repeating itself

unless memory/context is a separate open source layer of the agent stack, we're going to have to normalize the agent versions of Chrome, iOS, and Windows right?

Bragging about how much software you’re shipping with AI is like holding down the shutter button and bragging about how many photos you took.

@0interestrates one serves a functional purpose, the other should be making you feel something at a layer past the literal text

why do people (including me) have an aversion to AI writing but not as much to AI code? if a piece of text smells AI i stop reading it but i use things coded entirely with AI every day

verification seems to be the biggest bottleneck for enterprise agents?

@raflinens 100%. still working on this myself but I do think there's a common thread of strengths and weaknesses across all five types

@yb_effect Would you say the types are iterative, that over time in areas that are strengths/weaknesses it can become more intuitive and improve off each other?

A framework on how you can think about agency in different parts of your life