Decode:High Throughput Agentic Engineering with Kun

right now I'm using grock 4.5... I like 4.5 better because Grock 4.5 is faster and more efficient

Current primary LLM model being used, explicitly stated preference

I use both pi and cloud code. I basically use cloud code for whenever I want to use an anthropic model

Secondary agent harness specifically for Anthropic models

it's a config file in the config directory called crew dispatch.json

Uses JSON for configuration files, explicitly shown

this is done by a custom PI extension that I built... it comes with a first mate repo... called com

Built custom extension for PI agent harness to hide tool call noise

First mate is the essence of my workflow... it is a open source project that I created

Explicitly mentioned as his core workflow orchestration tool, open source project he created for managing multiple AI agents

I use pi... I basically use cloud code for whenever I want to use an anthropic model... And I use pi for everything else

Primary AI agent harness used for most models, explicitly mentioned multiple times

this mini is my Mac Mini which is actually on my shelf right now. It's running headlessly... the mini has basically more hardware resources than my MacBook

Uses headless Mac Mini for running multiple agent sessions with more compute power

This current local machine is a MacBook. It has a little bit less power than what my mini has

Primary local development machine

we will start with a clean terminal... I just launch my agent

Heavy terminal-based workflow with custom CLI tools

I have a cursor... if the task is a simple bug fix... I use... cursor

AI coding assistant used for bug fixes, has subscription quota

ask the crew mates to give me that in lavish for me to review... it's a really nicely structured interactive HTML

Tool for generating interactive HTML reports/artifacts for reviewing agent work

You just just clone this repo and launch your agent

Uses Git for version control, GitHub link in description

Herder is basically my new T-Max. It's a multiplexer. It manages all my agent sessions

Explicitly mentioned as terminal multiplexer for managing agent sessions, replacement for tmux

it's running opus 5... for new feature development work... always use Fable

Used for complex tasks and UI work, mentioned in crew dispatch rules

I use this tool called quota axi. That's another project I built. It can give us this data where I can just easily monitor how my quota is doing

Custom tool he built for monitoring API quota across multiple LLM providers

I have this Discord channel... I can just add my first mate... this bot is actually relay relaying this message back to my actual first mate

Custom Discord bot integration for relaying messages to First Mate

you notice that I use voice input by default... voice input is just faster than typing

Uses voice input as primary interaction method with agents

if the task is a simple bug fix... then I use Luna I use Sonnet

Used for simple bug fixes according to dispatch rules

backpass is a tool that you can use to help you improve your agent... it will analyze all your agent sessions in the past

Tool for analyzing agent sessions to improve agent configurations

on X I also sometimes when people report problems... this at this mention will also get relayed to my first mate session

Custom X/Twitter bot integration for relaying mentions to First Mate

it's a really nicely structured interactive HTML... lavish interactive artifact

Lavish generates interactive HTML artifacts for reviews

Herder has the ability to manage all the agent sessions across multiple machines

Herder manages agent sessions across local MacBook and remote Mac Mini

the task requires generating images... I use codeex as the harness and then GPT 5.6 soul

Used specifically for image generation tasks (likely meant GPT-4o, not '5.6 soul')

for new feature development work on this iOS app that I have... This app is actually already released in the app store

Has iOS apps in App Store (Edis Wallet, Ship app)

this whole world is procedurally generated. So it's always different

Fly with Me game uses procedural generation for world creation

I'm in the first mate directory... clone this repo and launch your agent

First Mate is likely Python-based given the 'py' agent harness and typical AI tooling stack

if the task is a simple bug fix... then I use Luna

Used for simple bug fixes in dispatch rules, Luna likely refers to Gemini

codeex has a native tool for generating images because openai has the image generation model

Uses OpenAI's image generation through Codeex for image tasks

when the task is technical product design architecture... I use fable I use Kim K3 I use Astra

Used for complex planning work, Astra likely refers to DeepSeek

this bot is actually relay relaying this message back to my actual first mate here running this local session

Real-time message relay from Discord/X to local agent suggests WebSocket or similar real-time protocol