Somil

Game design explains the shift in how agents should work better than most AI takes right now. We’re moving past heavy steering. The parallels are clearer than people realize.

another banger by colussus, but who's surprised colossus.com/article/sane-as…

very powerful anecdote from 1998 on the growing power and meaning of online communities. i read more about humans' relationships with technology these days as ai gets more intertwined with our lives a story about a tree raphkoster.com/games/essays/…

if all personal agents can bring pricing power back to consumers like grokbot just did, instinct's 30B valuation may just be worth it shoutout grokbot for helping me find @spaceship for buying domains

Religious fervor is back

running it back

Wake up, new anthropic s1 dropped sec.gov/Archives/edgar/data/…

unsurprisingly, replit argues that deferring to a model's native ability to spawn subagents with limited steering is optimal logical for a non-lab to argue this point since labs are incentivized to select their own models for subtasks rather than the MOST cost effective model out there vendors whose value adds include being model agnostic (replit, cognition) should aim to take advantage of this limitation great piece by the replit team breaking that down here: replit.com/blog/free-the-mod…

over the past 2 days, I've been understand why ramp's technical blogs hit so well selfishly want to apply these principles to my own writing at their core, ramp's writings: - feel like they were written for their internal teams - are honest - anchor their pieces around "what was understood" and 1) what they did differently, and 2) why will be stealing this formula for my own writing, thanks Ramp

another installment of forward deployed chips poker night bringing together NYC ai future founders, cohosted with @Work_Bench and the great @ProbyShandilya shoutout the awesome people from @DecagonAI,@OpenAI,@percepta,@SierraPlatform, and @DistSpectrum some notes from our convo: - personal agents are actually being used, the best one is still disputed - software factories don't seem totally real just yet - harnesses still matter - distributed systems are underserved by ai agents

a couple of reflections on the new code review process from the Claude code team at Anthropic: - they are playing a different game, I counted 5-10+ separate agent runs per PR which eats up so much context only a lab could bankroll - they've created incentives that promote ownership, the approver for a PR is responsible for any bugs or issues in production - review agents require quite a bit of restrictions (max # of turns, classifying bug severity, etc.) - standardizing the review process is still an open design problem - PREDICTION: there's going to be a more efficient way to see a snapshot of your codebase than having Claude generate a fresh artifact or summary each time - there's going to be (or maybe already is) an open source repo that provides a list of model-agnostic best practices that will be widely adopted claude.dev/blog/solving-code…

Really good quick snapshot on why salesforce acquired braze

now that agents do most (all) of coding, my apis should be mcps, and I needed a way to convert for my personal codebase but I wanted control over the conversion and scalable permissions So I built agentifyapi. I built it all with @bot and am blown away it does a few things: - every tool starts disabled - nothing goes live until I approve it - changes show up as diffs, with one-command rollback - Emails, phone numbers and card numbers get redacted Free and open source. check pip install agentifyapi github.com/SomilAgg123/agent…

insanely good piece from @dexhorthy at @humanlayer_dev on why software factories (still) fail things dex calls out as necessary: - design docs - system design visualizations - call-stack trees and file-tree diffs - vertical slices in shipping with personal codebases, ive seen the need for the same level of review as part of the software factory what's cool is i've seen you can embed these visualizations into each agent's run in your agent.md when the PR scope is big enough, so it's pretty trivial to implement this kind of visibility github.com/humanlayer/advanc…

sitting on the same side in a negotiation is really hard when your incentives don't reward it, but if you can find culture like early facebook, it makes it that much easier super important to be at a place with that culture

After using 4 different personal agents over the past week (instinct, muse, grokbot, dots), my ranking is: 1. Grokbot by a decent margin 2. Muse by a decent margin 3. Instinct 4. Dots