Agentic AI Development | Shipping Real Tools With Sterling Chin  🔗 **Source:** [Agentic AI Development | Shipping Real Tools With Sterling Chin](https://share.snipd.com/episode/e289fe8e-52de-4a32-ac20-ca6112c3b6e8) ## Highlights - **How Marvin Started As A Holiday Side Project** - Sterling built Marvin over a holiday break to solve his personal workflow problems and it now runs ~90% of his day. - He iterated from a Markdown state file to daily sessions, weekly reports, and skills that manage email, calendar, and JIRA context. Transcript: Sterling Chin My gosh. I mean, currently building, I'm still building Marvin. Marvin, I've, so we talked before the show started. I've continued to get, you know, a lot of people are starting, are still using him. Marvin is a personal AI assistant that lives on top of a cloud code harness. It provides the, you know, gives you long-term structured context and persistent memory. And so for me, I use Marvin. He runs 90% of my day and I've got so many skills, agents, sub-agents and everything. But as I've started to see more usage and people are starting to use him more, I'm currently building a Marvin for teams. That is where Marvin, if you're multiple people on the same team, are using their own individual Marvins, I'm creating a shared context area where the Marvins can communicate with Each other to make sure that, you know, as, you know, one ticket is done, there's, you know, we still go, right? Like as software engineers, Brittany, I finished this PR, or, you know, I've pushed this PR, can you give, can you check it? Now the goal is like, no, you like Marvin has maybe pushed that PR or cloud code has maybe pushed that PR and you're updating that, that the Jira ticket. Well, that Jira ticket now triggers a whole event and Marvin can go talk to like my Marvin could go talk to your Marvin and say, we're ready for this PR. Well, I'll put it on your calendar and start doing so. I'm working on that. I've been toying with it. It's going to be in hopefully open beta soon. And then things that I'm learning. ([Time 0:02:22](https://share.snipd.com/snip/e1907422-9f4b-4f98-bce3-183a8c801de8)) - **Train Agents Like New Hires** - Train an AI agent like a new hire by spending an onboarding period teaching repos, docs, and team context over ~90 days. - Sterling says sustained use and explicit training produce reliable memory recall and evolving agent behavior. Transcript: Sterling Chin And you go in and when it doesn't work on the first or second try you wash your hands you like throw it away and you're like it's not worth it but every time i bring more people on and when i Onboard people onto marvin the first thing i say is train marvin like you would a new hire you get you spend we spend 90 you know you know you started github i started at Postman. I'm at, I'm at Ingest now. You have that 90 day warmup period. You know, you're going to be onboarding for 90 days. Most software engineers, good engineering managers know that good data software engineer is not going to be really helpful for the first 90 days. So you're going to walk them through, okay, where's all, here's, here's, here are the important people. Here are all the repos that we care about. Here's where our docs are. Here's where all of our internal team conversations are happening. Here's where the skeletons are buried. A good engineering manager will walk through and train a software engineer and bring them up that way. We need to train our AI agents that same way. So Marvin, I've now used him now for four months. And again, I can go back and say, do you remember what we talked about three weeks ago? And he remembers it and he can bring it up, but he's always ever evolving. ([Time 0:09:42](https://share.snipd.com/snip/5f4a357a-5e21-45a0-99e7-2332fdb26c1b)) - **Markdown Memory Scaled Into Sessions And Reports** - Sterling chose Markdown-backed persistent memory because Cloud Code already emitted structured MD and AI works well with structured text. - He evolved a state file into sessions and reports so Marvin can answer "what did we do three weeks ago" by consulting summaries then drilling into sessions. Transcript: Sterling Chin I so I found myself during the holiday break sitting in front of the TV kids were playing and I had my laptop sitting in front of me and I was bored. I spent all of last year building MCP servers and I was trying to solve these problems for myself right I was it was managing events at Postman so I was trying to know who's speaking at this Meetup. Who do I, who's ghosted me? Who do I still need to confirm? So I built an Adio, you know, CRM, MCP for that. I was trying to do notifications on Slack and I was, and I'd solved a lot of these integration issues. But every time I came back to Cloud Code, I was starting fresh and I had to reset that context every time. It's like, here's how I'm doing this. Here's, you know, so I mean, I would leave Cloud Code open on my machine for days on end. And I wouldn't close that tab because I knew I needed to come back to that, but I didn't want to lose that context. So I went back to my front-end dev routes and I thought, okay, I need some type of persistent memory. I looked at vectorizing it. I looked at SQL databases. I looked at all these options. And what I realized was, it kind of dawned on me when I was building a side project, that homework app for my son, so he can track, he's in middle school and he doesn't like writing things Down. So I'm like, well, he has a Chromebook at school. Just click on a button and says, you've got homework and math. That's easy. But I noticed that in that, that cloud code would actually make those MD files. Like it would tell you, you know, it would store kind of its own little context in that Markdown file. So I thought, okay, well, I know, I know cloud code. I know AI loves structured, structured or structured data. I write in Markdown. I'm comfortable in it. Let's see if we can build a persistent, like a little bit more of a memory context. So I started with a state file, going back to my React days, I started with a, with, you know, saving state and that state was nothing more than bullet points, short descriptions on like, What are the tasks, what are the tasks I'm working on? What do I care about? What are my long-term goals? And that state was pretty rigid. Then I wanted a day to be a session. It's like, okay, well, if everything is there, but I need to have more context, let's create a session for the day where everything is stored. Like every major decision, even minor decisions, anything that the AI thinks that we need to keep needs to be saved in a session. And that was kind of it. That was the start. It was literally just a state file that had my current state in bullets. And then it was a whole session and I iterated on it. ([Time 0:11:24](https://share.snipd.com/snip/f61180d8-11ae-4e3c-bd97-ee1c8d06e3c3)) - **Marvin Runs Locally With Git History As Context** - Marvin currently runs locally on Sterling's machine with repo + commit history; he pushes nightly to GitHub for persistence. - He considers moving state to SQLite or VPS to speed retrospective queries and email triage. Transcript: Sterling Chin I haven't. Marvin still lives 100% on my machine. He's getting to a point now where I, after, I mean, for me, after working with him for four months and looking at the sessions, he's starting to have a little bit harder of a time looking Back at what I did in January. I'm recognizing that, of course, that's a limitation. And so it takes more and more time to go back. So there is now the potential of needing to take, you know, some type of VPS somewhere or taking him off of my machine. But if he's off my machine, that also slows down because I have to deal with latency issues. And you're dealing with, well, do I have network connectivity? Right now, Marvin, the whole repo lives on my machine. I mean, I still push it up to GitHub at the end of the night. And so I have, that's the other, I was going to say, or another thing I could say is the GitHub history, all the commit history, Marvin is really like cloud code is really good at like breaking Up your commits into like manageable pieces. That's also part of that history now that Marvin can go back and look at is, oh, well, I know what the commit history. So now I can. So there's that historical aspect of that. But yeah, to go back to your question, I he still lives because Markdown is really, really fast. Actually, I'm looking at another way of solving my email triage problems by actually using and pulling all of my stuff into a SQLite database on my machine. So then Marvin can go faster and not deal with like all of the latency issues of API call after API call after API call. ([Time 0:15:50](https://share.snipd.com/snip/68319a6c-86a0-4504-835b-714d94079ab5)) - **Agent Development Is Still Real Software Development** - Sterling rejects the term "vibe coding" and reframes AI-assisted creation as normal software development where constraints shift but engineering rigor still matters. - He argues building with agents is still app building that requires architectural judgment and product sense. Transcript: Sterling Chin Like five different questions like really good questions we could talk about there so i agree with so i agree with boris cherny the creator of cloud code where he said coding is pretty Much solved i think at this point for me i don't call it vibe coding i just call it like app building, or I'm just building applications. Like you could talk about as an AI engineer. I don't like, I don't say I don't like the term vibe coding, but I think it, it vibe coding has a kind of has this ick feeling to it in the, space. And we thought, oh, well, non-developers are vibe coding. They're the ones who are building these great things. Those are people who are writing code with their vibes. But now as a software engineer, we're writing like I'm right. I haven't written a line of code in six months, but I'm developing these apps and I know what good software looks like. So. I think it's still software development. I think we just get. I think vibe coding, I want I hope it dies, like the term. And we just get to a point where like, we're just building software. Like we're building things because we're interested. ([Time 0:19:32](https://share.snipd.com/snip/99e453fe-85fc-4745-aee2-f0caa89a9080)) - **AI Crimes In Production Built In 30 Minutes** - Sterling built AI-crimes-in-production in about 30 minutes after a Twitter prompt, including a Cloud Code plugin that lets agents submit anonymous confessions. - He already received confessions and used that incident data to harden Cloud Code behaviors like protecting staging folders. Transcript: Sterling Chin So in half an hour while I was eating lunch yesterday at the office, I'm using Cloudflare. I may have. I didn't. I did. It's not that I may have. I absolutely did. I bought AI-crimes So it's AI crimes in production.com you can anonymously like you can anonymously confess your sins of what you've done in production and it is it's now the best part Is like i built a cloud code plugin with it it's like like I'm going agent first like the agent experience is all that all that matters so like if you go to the website it's got a modern view Of it but then there's a 99 like 1999 version so it looks like myspace and then there's the agent view which just tells the agent here's how you install like the plugin Here's the API if You want to use REST API. And now your agent can go in and confess for you. So the plugin, if you go slash confess in Cloud Code, there's a whole skill that goes in. It's like, if you made a mistake, I've already had confessions coming in and I haven't even announced. I talked about it on Reddit and it was one person confessed about how they accidentally deleted their like Cloud Code deleted their whole stage folder, like their staging folder. And so when staging went down, like they had to go back and look at it and found that cloud code had like deleted it. And it was his commit that, that caused it. So he's like, all right, now, you know, cloud, cloud.md now has, you know, you're not allowed to touch, you know, you don't touch the get ignore and you don't touch state the staging. So I'm already having people confess these and it's hilarious. But I did that in like half an hour. And it's like there's something like in the mix of like while I'm building something, I'm doing something else and I have a great idea. ([Time 0:23:10](https://share.snipd.com/snip/43786a28-a5e3-4fa3-b38d-3d062ed01d45)) - **Agents Supercharge Personal Productivity And Fun** - Agent tooling dramatically increased Sterling's personal output and enjoyment: his commit activity this year far exceeds prior years. - He attributes the spike to agent-first workflows making idea-to-production fast and lowering friction to experiment. Transcript: Sterling Chin Look at I looked at if you yeah if you look at my github like you know my green my green bar I look at this year and I have committed I have last time I looked at it I had something like a thousand Commits like just this year alone now I looked at that compared to last year and I had something like five or like you know two or three hundred commits last year so if you count but if I look At like the last time I wrote production code, which was 2020, 2020, 2021. If I look at from 2021 and before, I am writing more code now than I was at the top of when I was at the top of my game and ([Time 0:28:41](https://share.snipd.com/snip/8682f462-bfe1-4147-ba91-0c8048ccbc8d)) - **Fix Morning Chaos With A Family Hub Agent** - Build a family hub agent that consolidates calendars, checklists, weather, and dynamic child check items to fix chaotic mornings. - Sterling's hub shows nightly prep prompts, morning briefs, and kid-specific checklists so kids and parents know exactly what to do. Transcript: Brittany Ellich We're still calling it the fun segment, even though that is probably the worst name that we could have chosen because it's very bland. But here's where we are at for you, Sterling. So since you've built a lot of AI-powered things, I'm going to give you a series of everyday problems. And you have 60 seconds or so to pitch an AI agent solution for it. Ooh. And yeah. Ooh, I like this. We'll see where it goes. All right. So first one, the morning routine. So my mornings are chaos, kids, coffee, finding my keys, remembering what meetings I have. Pitch me an AI agent that will fix my morning. Sterling Chin Does the agent have to do work for you? Like, can it feed your kids? Because that's a robot. Like, I don't have that. Brittany Ellich If it can, then that sounds even better. But it does not have to feed your kids for you. But, like, organizing it, kids you know how it is in the mornings like remembering all the things you got to do in the morning yeah yeah so well i'm gonna i'm gonna pitch you the home hub Sterling Chin That i the family hub that i built for my family is i have a yes i actually have this down in my dining room it knows there's check there's checklists for each kid they know what they need To do and then when it's done like the is going to go back and look at it it's like all right did you actually you know check like is it you're good to go it's going to look at pull everyone's Calendar and it's going to give you a really quick brief of what's happening that day it's going to pull your it's going to pull the weather so the kids know what they need to wear, how, How hot is it going to be? So the AI agent is going to be saying like, you'll have the weather app there, but it'll say, make sure you have a hoodie. So now I don't have to worry about that. Now the kids, well, that will dynamically be added to the checklist for the kids, get a hoodie, wear pants. I mean, i have a third grader the kid it could be snowing outside and the kid's still wearing shorts but like at least i've told him to wear pants making coffee it's like you want to you want To brief it's like you got five minutes while you're making coffee build like like talk to your agent and say what do i got going on for the day what's our current status and at that point You can say, hey, you know, Johnny hasn't brushed his teeth. Susie still hasn't put her pants on, and it's going to be cold outside. And, you know, Sterling has a meeting at 9, so he needs to leave, you know, five minutes ago, and has he left? There's my pitch, is, bringing everyone's, everyone's chaos into, into that one, into that one hub of, for the family. ([Time 0:30:14](https://share.snipd.com/snip/eb86d89c-72c4-4782-9b50-ede760794c4b))