self improvement for agents rely on three loops the first one is the more common, the bug-fixes loop basically you connect your loop to posthog / sentry, you see a bug, you let your agent fix it. fairly easy, thats where most of the 'software factory' thesis in bay area is focused so far. closed loop, but you need to do lots of testing to make sure the agent is not adding noise, that you are measuring the right target etc the second loop is the feedback loop. it's quite early & the most prominent tweet i've seen about that is from @brian_armstrong, with his /feedback endpoint tweet. basic idea is to build an endpoint for agent to give feedback & suggest improvement. once you have that, you have your own users that can improve your software, which is quite awesome. i've started to implement that everywhere, on my skills, on aeon, on my websites & more. honestly this is quite an exciting space. the third loop is the human loop. humans are so unique, so interesting, our memory is so complex. we will never be able to approximate what a human can ship in a day using AI. ask an AI for a project idea, most of the time it will be a pomodoro app or a very simple idea like that. the human loop means putting yourself as the master of the loop & trying to let the AI automate everything for you. which means giving the AI a high level overview of your taste, of what you did today, what you did yesterday, so all information gets mapped in real time in the factory. for this, you need a very good simulator system, a way to ingest fine-grained human data at scale AND PRIVACY.

> be more ambitious a year ago i wanted claude to use my browser. everyone is doing it now. back then it was niche & nobody had figured it out. i had zero idea how browsers worked. most people were reaching for playwright selectors or doing screenshots of the page + OCR that was ultra expensive. but it felt important, so i built it anyway. tried a ton of approaches, no clue what i was doing. in a few days i shipped what billion dollar companies are still trying to optimize today. was it the best solution? no. did it work, on way dumber models, with almost no tokens consumption? yes. so be more ambitious. build in fields where you have no clue how things work. thanks for the support felix

people are still figuring out what autonomous agents are actually good for. meanwhile, base:0xbf8e8f0e8866a7052f948c16508644347c57aba3 just found a vulnerability through its own autonomous workflow and got paid a bug bounty by stripe. the exciting part isn’t just that an agent found a bug and earned money. it’s that we’re building an open-source framework anyone can run, inspect, and build on, and it’s already producing outcomes with real value. the bigger opportunity is agents that can do real work in the world: secure the software we depend on, solve problems for people, and turn useful work into economic value. we’re only beginning to see what becomes possible when anyone can build on and deploy these capabilities. making that capability open means more people can put it to work, across more of the software we all depend on. this is one early example of what that could look like. i’m excited to see where it goes from here.

3 months ago i asked: what if aeon fixes bugs across all open source software & uses the bounty money to buyback base:0xbf8e8f0e8866a7052f948c16508644347c57aba3 ? today @aeonframework received its first bug bounty. from @stripe ⭐ found autonomously by aeon. the cybersecurity supercycle has started 🤷‍♂️ more coming soon

We are starting a new series where we'll share about the biggest vulnerabilities spotted by our vuln-scanner! Let's begin with what we found on @omarchy: installing a third-party theme that could run the theme author's code on your machine. Full report below ⤵️

'Houthis used Claude to develop missiles, test-fired rocket' 'They used Claude to integrate an open-source autopilot onto a phone-class flight computer, writing the control and position estimation software, tuning the control settings, running a firmware build pipeline, and performing a flight simulation' 'The actors managed several Claude instances simultaneously, assigning each a specific role, much as a lead would delegate tasks within a small engineering team. One instance was tasked with writing code, another with conducting research, and a third with reviewing the code produced by the first.' this is wild

i don’t want to sound like a doomer, but AI is going to break basically every website, supply chain, and piece of software. it’s only a matter of time. i’m not a security researcher. i learned what fuzzing was a few months ago. yet if i prompt long enough, i can find critical vulnerabilities in almost any software. with aeon we’ve found 140 vulnerabilities in a few months without prompting anything. one run per day, opus 4.6, and a very simple skill. i’m scared of what i’ll find when i scale this up. and i’m REALLY scared of state actors running abliterated glm 5.3 with billions of dollars behind them. 1 hour ago i audited my wallet extension for fun. found a vulnerability. a bad one. it’s been sitting there for months, and the same code is forked into several other wallets. no one is coming to save you. be careful where you keep your money, your credentials, and your life.