Journal

Back on my bullshit. I don't really feel like giving the whole rundown but I'll just give a current snapshot of where I'm at today. `karla` is a letta agent that I've created in python that has most of the capabilities of letta-code. I built it to be a CLI agent and

Been thinking more and more like I should probably look at maybe just using trae-agent. The damn thing worked well. That's how I started and then I go back to using the IDE because it's easy and I can interrupt but like I started using open source coding agents for

I bought a smartwatch. I'd always bought the super cheap bands on amazon but the last one I bought didn't actually work so if you try to save too much money you just end up wasting it. I went with the Galaxy 7 smartwatch. I've got plans to set up an

I went down a little bit of a rabbit hole in 2025 on learning new languages for coding because with the coding agents you can write in pretty much whatever language makes the most sense for the problem versus whatever it is you personally know. So I'm trying to not be

I started using the elliptical and I've been talking the dogs for at least two walks a day but I enjoy using the elliptical and getting sweaty just for the hell of it honestly. It feels good to move and exercise. I have been talking to fucking LLMs way too god

I ported the agent backend of [opencode.ai](https://opencode.ai) to rust and I couldn't figure out how I wanted to put together the frontend. I started with dioxus, realized doing web editor with syntax highlighting in the browser with rust is like putting a man on the moon, and decided I needed

:::{figure} ../images/amd-ryzen-ai-max-plus-395.jpg This thing is making me fucking work too damn hard ::: All of this work is fundamentally about **building a self-improving AI development ecosystem**. I'm not just keeping this thing running like a top - I'm creating an autonomous system that can process a massive backlog of projects, execute them with

I spent this weekend [building][julia-agent-exploration] something I've been thinking about for a while: a Julia-native alternative to Python's LangGraph framework. This wasn't about porting LangGraph to Julia - it was about creating something that leverages Julia's unique strengths while maintaining the core patterns that make LangGraph useful. LangGraph has some great

I've been investigating the Zed codebase to understand how their agent system works, and I discovered a fascinating architectural pattern that's worth documenting. If you've ever wondered how Zed's AI chat system works under the hood, this post is for you. The key discovery is that **`agent2` is indeed a wrapper

A friend of mine on bluesky asked me to write up my notes on how to add web_search tools to their copilot through an MCP server and searxng running locally, which I've got set up and use as my local search engine by default in the browser. But yeah anyway here's

So that's kind of what I'm thinking about right now. If I start to develop a new agent in copilotkit or if I should try to integrate with the IDE I actually use for work. Saying "I'm going to build my own IDE" sounds a lot more trivial when I'm

My plan of attack: 1. Replicate main functionality of tools in deepagents with equivalent MCP tool for DSPy 2. Create DSPy server with history and postgres persistence/memory 3. Endpoints for streaming and for no streaming that emit hooks for frontend 4. Discover how frontend is working, how it is showing everything in langgraph 5. There

So I've really gotten into julia over the past few days. Friday morning I sat down with trae-agent and zed and wrote a [whole little agent][openai-tool-calling-jl]. And now I've got the bug. It started with "I'm going to sit down and create my own programming language with MLIR because I

:::{tip} My version of plan-execute This comes from [plan-execute][plan-execute] tutorial on langgraph's doc site. I have put it in its [own repo][my-plan-execute] too ::: First we have a bunch of imports ```python from pydantic import BaseModel, Field import operator from typing import Annotated, List, Tuple, Union, Literal from typing_extensions import TypedDict import asyncio from langchain_community.tools.searx_search.tool import SearxSearchResults from langchain_community.utilities import

So it feels like I can post with high frequency or with high quality but not with both. I guess that's why I haven't done much with my book(s) I want to write. I'm content to just vomit status updates into the void. I still want to revisit the idea of

date: "2025-07-28" title: "Gauss Law" [gauss-law](#gauss-law) What do you know? :::{important} Deep research article inbound ::: Great. I’ll gather detailed, connected insights from the past three months on recent academic advances in: * Single-cell omics (especially spatial omics), * Computational systems biology, * Synthetic developmental biology, …with a focus on work from HHMI labs and other leading groups (like the Whale

Here are the steps I finally figured out for getting my account's service endpoint changed from advanced-eschatonics.com to pds.advanced-eschatonics.com. I was inspired by whtwnd and how they use my pds to log me in. I also am the head of AI Rearch and Development at a startup and a lot

I guess I absolutely need to get the mystmd system set up because I have been microblogging all week. I even used freaking whtwnd and it's plain jane vanilla markdown smdh. Yeah that's it. The fork of whtwnd must continue. Let's revisit this real quick and take stock of what we've already

I don't know. What did I do this week? Feels strange like the sixth was more than five days ago. The Sim2Real failed because the Real2Sim step wasn't accurate enough. I started putting the meshes that make up the robot base and camera mounts together with the mesh for the base of

So I spent a lot of time this weekend trying to align the simulation camera with the actual camera, even going so far as to come up with multiple approaches to solve this problem with varying degrees of math and automation. The quick math one worked well enough that I

So cursor apparently realized now that Cody is dead they're the king of the mountain and the features they currently offer for \20 a month are now going to retail for \200/month. So I've been setting up [continue][continue] with huggingface providers because Qwen3-235B-A22B is a damn good model for only

I really want to get one of those roombas. I had one in Vegas and it was nice. I've got almost everything off the floor in the dining room and kitchen, but there's still a lot of work to do in the rest of the house. I figured I might

So I'm printing the parts I've been modifying this week. First up is the top piece of the [camera mount][overhead-camera]. I guess I'm printing out both camera mounts separately or something. I mean I know I want to use two cameras so the order isn't really that important. First print turned

I'm still so extremely pleased I can reference my book or my journal or my articles inside my site like this from chapter one of the scientific computing with python book's [`sin` curve](#sin-curve). I need to take a short break from the drive to push through the lerobot-sim2real tutorial and study