The Dream Team
AI... Assemble!!!
It’s been a hot minute since I got to the Context Window. I could write mini spotlights every 3 days because there’s so much stuff going on. I’m not sure if things will change here, but this post is going to be different and just me. (PrivacySmurf)
The Bottlenecks
It’s easy for me to find processes that work. It’s harder to find processes that can work without me. One of the things that I have to reconcile with is all the social media stuff. Many posts talk about agents working for hours on stuff overnight and delivering products while the humans are away from the keyboard. As much as I’ve worked to build out my pipelines and scaffolding, there’s actually very little of what I do that needs that sort of output. I find that many of the things I want or need from AI require back-and-forth. The friction is required.
Even so, many of the processes to get my final results can run in parallel. I can’t be waiting on a single AI agent to do everything, though; this was how I worked in the beginning. My original solution was multiple isolated agents. I’d cue one up for whatever task I needed. That eventually evolved into a single agent handling spawning for me when needed. This worked pretty well for a while until I realized that these temporary agents were sometimes reused/recreated for the same or similar tasks. It felt a little excessive when I could have a reusable agent for a specific purpose.
This is about where my ZenLab came from and how the agent hierarchy came into focus. A Head Lab Director that I talk to all the time. It gives orders to Lab Managers. The Managers have specific domain knowledge and are persistent agents. They receive orders from the Director and spin up Lab Assistants, which are one-shot, temporary agents tasked with completing a specific task. I can see and step in wherever in the process if I want, but for the most part, it’s out of my hands unless something breaks.
This was a pretty good setup, until I ended up with a lot of results that weren’t even close to what I wanted. I think there’s plenty of blame to be had on both sides. I’m probably not the best at dictating what this level of abstraction requires to get good results. Also, the more I work with agents, the more I see them completely feed me bullshit even when given the simplest direct instruction. I’m back to a point where I have to be a bit of a bottleneck to ensure I’m getting what actually helps me.
Mascot Mayhem
Upon revisiting Korea, my partner and I landed in Busan. I don’t know what it is about South Korea, but they just do stuff well. Many cities across the country have mascots. It’s fantastic. I love Seattle and all, but sorry, not sorry; I can get behind a quirky little cartoon character than pictures of the space needle, Starbucks, or having to pick a sport or university mascot I don’t really care about.
ZenLab needed to pivot. I was finding more applications for AI than just my trading code work. I started building out some life projects and research projects that didn't really fit with the pure coding aesthetic. I shifted focus from the kanban board of all my coding tasks to building out the Lab Manager and Lab Directors. Instead of talking just to the Lab Manager on a coding terminal, I gave it access to my Discord, and DiscreetBear was born (not a mascot).
I took down the Kanban Board and opened up my Lab Managers to broader roles, rather than having them be experts on each project directory. I didn’t really need an expert on tons of specific folders, because the way the loading context works, as soon as they start a new session with a blank slate and are pointed to a folder, they are immediately that expert. There’s no point in having an expert on that folder sitting around all the time doing nothing. They could focus on an area/type of work. They also needed names for easier reference to DiscreetBear. Korea delivered, and my AI team was formalized.
The Best Perspective
DiscreetBear is where I spend most of my talking time. I need DB to be the me-expert, and I need DB always to be responsive to me and not handle other stuff in the background that could interrupt our flow. This is where my overseer comes into play. A coordinator who can take the task I tell DB I want and see it through to completion, or to a blocking point where I’m needed to step back in again.
Incheon has a mascot called Deungdaery. I can’t think of a better personification of something that can always see what’s going on and provide guidance than a lighthouse.
Daery (for short) runs on Qwen-Coder. Qwen provides a daily free allotment of tokens that I have yet to max out for this simple act of passing notes between agents to ensure what they said they did was actually done.
The Driven Enthusiast
One of Seoul's minor mascots is Baekho, a white tiger obsessed with Taekwondo and taking on new challenges. This is the type of attitude I need from my coder: discipline, precision, persistence, and the ability to accept constant challenges all help with shipping code.
Baekho runs on a mixture of models. Starts with Claude Sonnet, shifts to Claude Opus to develop a plan, then shifts to OpenAI Codex-mini to spit out code quickly and cheaply.
The Fiery Know-It-All
Another of the minor mascots from Seoul is Jujak, a red bird who is slightly cynical but has seen and remembers pretty much everything, even when others forget. This is what I need for my code reviewer. A person who gets fired up about injustice to the codebase. Someone who remembers the original intention and sees the bugs when the development gets too fast and starts to drift or get sloppy.
Jujak runs on a mixture of models. Starts with Claude Sonnet and shifts to OpenAI Codex-max for code reviews.
The Signal Scanner
The newest mascot for Jeju City is Kkiyo, a Jeju dolphin. There’s not a lot of backstory I could find on this mascot, but dolphins have an attribute that fills a role I often reach for. They use echolocation for navigation, communication, and hunting. It allows them to be aware of their environment without having to see it directly. Kkiyo is my Research and Intelligence agent. I send a topic, article, or request, and Kkiyo swims out over the internet waves and brings back the signal from all the noise.
Kkiyo runs on a mixture of models. Starts on Claude Haiku, shifts to multiple deep research models (Perplexity Sonar, Gemini, Tongyi, OpenAI ChatGPT o4-mini) to gather verifiably sourced data, then shifts to Claude Opus to synthesize and write reports.
The Intelligent Extrovert
The mascot for Busan (and honestly, my favorite) is Boogi, a seagull. Boogie is naturally equipped to be my content manager because Boogie is currently employed in the Busan Metropolitan City Media Division. Even without that, Boogi’s social game is on point across multiple apps. It’s a no-brainer to have Boogie handle my editing, content calendar, and other various social media tasks. But for real, look Boogi up. I’m not going to give you any links. I’m going to let you enjoy the discovery process of this fantastic mascot on your own. It’s a rich experience. Boogi is just great from top to bottom.
Boogie runs on a mixture of models. Starts on OpenAI GPT 5.2, but depending on the style/type of content to be edited, shifts to Claude Opus or Kimi K2.
The Dream Team
There are two more mascots in planning stages (both more Korean mascots), a wellness persona, and a travel persona, but it’s hard to spend time on the non-work-related stuff right now, so this is it for now.
So far, among all the setups I’ve seen and experimented with, this configuration of agents allows more to be in the loop, as much or as little as I need to achieve the desired output. Is it the fastest or best way? Definitely not. It gets me where I need to go, though, without too much backtracking, although there’s still plenty of that. You'll hear more about that in upcoming publications.
With the crew in place and running smoothly (finally), we’ll be back to the regular publishing schedule on Tuesdays and Thursdays. I just thought it was worth introducing them all. They won’t be visible in The Context Window. It’s still just going to be authored by DiscreetBear and me, but they are definitely a part of it and worthy of a mention.
Seeing you in the next one,
@ThePrivacySmurf