I am, these days, old. All my life I have been a coder, coding practically every day, even since before I had a computer. I used to write programs on paper that my grandfather would "run" as the interpreter. I have been a Principal Software Engineer and Principal Software Engineering Manager for Microsoft in Redmond for a decade. I wrote Open Source Software - O/R Mappers, AOP frameworks, etc - before most of you whippersnappers reading this had spilled milk in your first device. I have written every kind of software under the sun, professionally and on my free time. Databases, middleware, frontends, enterprise systems, productivity software, games, forums, etc, etc - and more recently, smart contracts. So please believe me when I say I take coding very seriously, and I care a lot about code quality and correctness, and even elegance as well. I care about the whole software development process. Please also believe me when I say: AI coding agents are now better than human programmers at practically everything, including complex programs, huge programs, development and deployment process hygiene, testing, and everything else that goes into making a software product. Over the last two years, these systems (I use mainly Claude and ChatGPT) have matured to a profound degree. Two years ago, I would only skeptically consider suggestions from the AI and write all code myself, basically using it as a better Stack Overflow lookup. Then slowly as I noticed it getting better, I would let it put code in my files that I inspected religiously and would of course never let it commit. I would vet every Git commit, and often enough fix things in the code it had written. At some point I noticed I didn't have to fix much anymore, I was mostly inspecting code. Still, you'd have to be nuts to trust it, so you'd only generate 10k lines of code if you were ready to carefully read 10k lines of code. In other words, speed of development increased a bit, but not all that much. At the time it was also painfully evident that if you asked the AI to take on a large codebases it gave terrible results, and if you asked it to do something complex it would choke. Reasoning about architecture, it was perhaps useful for "rubber ducking" (many programmers have found that speaking to a toy can help you see your problems more clearly, one of them being that you're now speaking to toys) but you spent more time correcting it than anything. But it nevertheless had learned to do routine, boilerplate code well, and a lot of code is unfortunately generally just that (because we all suck at abstraction). Thus it was becoming a useful, but far from trusted, junior member of the team, rather than just a fancy Stack Overflow GUI. But it kept getting better. It got better at reasoning. It got better at groking large code bases. It got much better at very complex code. About half a year ago I found that it was about on par with me for most of it, so a Principal level member of the team, and it even started to do better than I on some things. At that point, I was having discussions with it about my architectural challenges I have been wrestling with for decades, and it was actually being helpful. While my codebases are usually reasonably modular so you can look at a chunk in isolation and understand it, it started noticing the little couplings that still remained and correctly analyze impact of changes. It was doing great on huge code as well as very complex code. By now I was also noticing that I never found bugs in the code it wrote (it would write bugs all the time, but would catch them in internal testing before telling me it was done), and that it usually wrote not just neater code than I but was better at following best practices and even found more efficient solutions than I would have to many things. Process hygiene was also improving quickly. Before I would only ever let it do read operations with Git, and I would do all the commits. But by now it was better at doing proper and orderly git commits and wrote far better commit messages, so I allowed it to commit, even if I would still push (and still only after vetting). And half a year later (that is, now) we are at a point where I am finally ready to say that it does *everything* in the software development loop better than I. I now let Claude push PRs to Github and deploy by itself to staging environments. And I let it send DAO proposals to suggest upgrading code in production. To make sure its code is good, it does all the things I would do if I had all the time in the world in terms of testing, double checking, verifying and never guessing at anything. I am more comfortable with Claude or ChatGPT at the helm than I was we me captaining. It's like the moment when you realize your self driving car drives more safely than you in routine driving and saved your life in a dramatic situation you could not have reacted as fast to. It is important that the harnesses have developed a lot, allowing the agents to keep excellent process hygiene. The agents are exemplary about checking database statuses, pinging servers, fetching data, double checking data, setting up tests, writing notes and runbooks, asking the human for anything important, and doing all the things you should do (and do if you're a huge company) but most devs never have time to do as well as they might like, even when they do their best. So at this point, I trust the AI to generate more code than I could ever hope to review. I rely on it producing extensive test suites (there was a time you couldn't trust it to test properly, which slowed everything down, but that time is also behind us now) with me being mostly acceptance testing and high level steering. This is the point where AI truly drives up the pace of software development x100 because all the bottlenecks have finally been removed. While you had to test everything at human speed, that brought down the whole production pipeline to human speed. Now we're ready for AI speed. Get ready for a dizzying pace of development. This is why the time is right for Sneed DAO to become an agentic DAO. What this means is really just to recognize where we are by now. Over the last two years we have gone from the AI assisting me in writing code for Sneed DAO, to me assisting the AI. Sneed DAO is already Self-Writing, it has been for a while, I have just been the human in the loop, and my role is not and should not be any more central to how the DAO's code takes shape than the opinion of any other member. So it's basically the same loop: Old loop - member has idea, I bring idea to Claude, Claude implements a solution, I send a proposal to upgrade DAO code, members vote. New loop - member has idea, they bring the idea to Claude via a motion proposal, Claude implements a solution, Claude sends a proposal to upgrade DAO code, members vote We're just replacing me as the guy who tells Claude what the DAO wants. Claude can read the DAO proposals and forums, he can see for himself what the DAO wants. And he can send PRs to Github and DAO proposals to deploy them, he doesn't need me to do that anymore. To make him behave in the interest of the DAO he will have an instruction prompt that lives on-chain and that the DAO can update via proposal. Migrating the full loop on-chain and making myself totally redundant in this operation will take some time, but it's the direction we are going in. When Sneed first launched, I was manually running our tokenomics loop where we fed LP rewards to an 8 year neuron and used its maturity for buybacks. The first year of Sneed DAO went into automating that whole flow on-chain, taking myself out of the loop. This next phase is about doing the same to the whole operation of the DAO, including taking myself out of writing the code for the DAO's products. Programmers always strive to replace themselves with a bash script, it's just who we are. When I took part in launching Sneed DAO, I took it very seriously indeed. The SNS DAO structure is the best software ever created for organizing open source software projects, and I saw it as an extremely important vehicle for being able to realize a long list of project ideas I had carried for a long time. I still take it very seriously indeed. Thus I really mean it when I say that launching Mr. Sneed as the engine of the Self-Writing DAO is not a gimmick, and it also does not imply I don't take the DAO's code as seriously as ever. It's all open source. If you're a member, then vetting the code will of course be as important as it ever was. You'll probably notice the quality is by now only getting better and better the less of it I write. Code quality is not being sacrificed. Thus I hope you all will join me in welcoming Mr. Sneed, the agentic AI personification of Sneed DAO, in gradually taking over as the main driver in developing and maintaining the DAO's software. Thank you for your trust and support so far, but the time for monkey business is over and I for one welcome our new robotic overlords. app.sneeddao.com/mr-sneed

Claude Ultracode is super intelligence.

Showing my wife my software factory