Chrispian

AI got us out here counting toes.

If you think using tools like claude code is agentic development I have really bad news for you.

Getting really good feedback on my AI rollout proposal for work. Getting to design an entire AI/LLM platform for a company this size is something I never dreamed I'd be doing when started playing with computers when I was a teenager. What a time to be alive.

I gave my agents messaging (1-1) and groups so they could coordinate complicated cross project, multistep tasks where timing and order of operations matter. Along with the other tools I've built it's enabled a much smoother workflow. I'm very close to the final abstraction layer.

It's so cool seeing how far my system has come. Reached a major milestone that's going to unlock a much more productive workflow and remove the bottleneck of trying to operate agents at my scale through CLI. And I don't just mean a GUI.

So excited. Finally to the point where I can really push the chat harness and I know I'm biased but this is really good. I'm very proud of how this is coming together. This is the last major hurdle, once this is stable I can release it for people to check out.

You merely adopted the cli. I was born in it, molded by it. I'm bringing amber and green chrome back. My first monitor ever was amber chrome.

Every time I see an anti ai post the complaint is usually about shitty human behavior followed by a sales pitch.

Great minds think alike. Here's my take on a chat scratchpad. there are other cards in the bottom chat drawer like a mini terminal. I can tell @aarondfrancis really uses AI heavily for us to both land on a similar solution. This is stupid levels of useful.

Clockwork is a task manager (think Jira) for agents. Tasks run through a queue and I can monitor each task in real time. This is not chat, this is automated, durable tasks that can run 24/7.

LongMemeval finished on the s track and it's very good news. There's a little bit of tweaking to do and we'll be even higher. Oracle is about to run and it looks like we are going to do really well there. If everything checks out, this is publish worthy.

LongMemEval on codename: Vanta will be done in the morning! Huge score jump now that we are at 70% done. I'm pretty excited to see the score in the morning.

I found a voicemail my Dad left me. It was really short, very typical Dad voicemail. And it broke me. I have complete Aphantasia so I can't imagine or recall sound, images, etc. I'm very happy I found this and so heartbroken every time I listen to it. I miss him so much.

I can't wait to show off this reactive interface. And I don't mean "React" reactive. I

The fact that we label things made with AI will be a joke in 5 years, tops. Nobody says "This post was edited with Microsoft Word".

Time to run the full LongMemEval on Codename: Vanta (my agentic memory app). The first run was very strong. Did 2 very minor tweaks and the smoke test results are already way higher than the first time. Hopefully will have the results later tonight. Watch out Milla Movovich!

High score? What's that mean? Did I break it?

Made with AI + Tailwind + Shadcn + Astro. Really happy with how this turned out. Still have a few polish passes to do but my new dev site will be live soon.

All the tools I've built lately are AI/Agent first. That's an important design decision and it's paying off for me so far. And since these are products for agents, I ask my agents for product feedback just like we would for users. This one thing alone has been super valuable.

I can see live updates in Clockwork (agent task manager/runner). Tasks are durable context objects and everything needed to work the task is in it. I have a Plan feature and an orchestrator picks up a plan and executes it in phases. Love how well this is working.

How my agent's plan a feature. Planner creates a plan, phases, sprint & tasks and then an orchestrator runs and makes sure each one goes smoothly. All through an automated task runner I built, this is the GUI so I can observe in real time. Phase 1 of this plan finished just now.

Seeing agent's spontaneously use the memory system I made it so damn cool.

Demo of the activity widget for my agent multiplexor (proxy/router/messaging/session management).

I have @opencode running through my local MCP proxy (Agent Mux) and one of the bonuses is that I get real time activity from every tool call. It also opens up PII protection and anything you could normally do with middleware. I get Agent to Agent messaging for free.

Just got my MCP proxy working locally and now it's just another provider to Opencode. Setup one MCP and now all MCP's are available. + I can do orchestration/messaging between sessions now. My agent multiplexor works with Claude and Open code! Just have to add more adapters now.

Tokens or commits?

The first run of the memory test went well. Found a couple of small issues and tweak those and the score went up a good bit on the samples. Now to re-run the full set and compare to the baseline. Still pretty confident it'll be a strong score. Full results should be tomorrow!

The first run of the memory test went well. Found a couple of small issues and tweak those and the score went up a good bit on the samples. Now to re-run the full set and compare to the baseline. Still pretty confident it'll be a strong score.

Very promising early results from memory benchmark on my system. Out of the box running against a published benchmark, zero tuning, and already 15-20 points above the embedding baseline. It'll be a few more hours for the full test to complete but this was the "hard" section.

Setup a test harness so I can test AI memory apps (LongMemEval). Running it on my own memory app and it's 5/5 so far on the golden examples. Not a large sample but happy it's off to a good start. It'll take a few hours to fully run (first run embeddings). I predict a high score.

Almost done with a major foundational update to my ai chat app. This closes the remaining plumbing gaps and everything from here is mostly behavior/business logic using this plumbing. So excited to finally have all this part done. Really proud of how this is coming together.

Just had the best validation of my work probably in my entire career. And I can't tell you about it. UGH! I'll be able to share in a month or two. So excited, I can't wait!

Does your agent/harness have a dedicated tool drawer? Keeps the chat clean but you don't lose tool call details if you need them.

Major progress on Nanite the last few weeks. This quick demo shows how agents & human can collaborate on session todos and plans. The user can mark items done, undone and send it back, drag/drop to sort, etc. This is just a sneak peak, you should see the entire workflow!

Finally down to just testing/docs for Nanite. Should not be much longer. Also secured domains for the next 2 apps I plan on releasing after this one.

If you play videos at 2x or more, it's much easier to spot AI produced videos. Just throwing that out there.

Anthropic using regex instead of their actual *language model* to detect negative sentiment in prompts.

People say agent orchestration when they mean "supports multiple providers/models". These two things aren't even similar. Really chaps my harness.

I'm a copyleft / anti-copyright type person. But you can't really be copyleft and also be mad when people take something open and make a closed version or don't give back (eg: WP vs WP Engine). Free means free or it means nothing. writings.hongminhee.org/2026/03/lega...

I really miss the days when you could buy a good domain for a project for face value. I worked with domain squatters in the dot com era and they ruined it for everyone then. Today it's developers hoarding domains for projects they'll never get to (myself included).

Got the plugin system for my AI chat app done. I'm really excited about this part. I've spent a lot of time making sure plugins are first class. All non-core features are plugin based and users can create their own plugins. Basic marketplace is also done!

Probably have about a week's worth of work left on my ai chat app. I know everyone's making these, but this one works exactly the way I want and I suspect there will be people who like the choices I made. I have to rename it because too many apps have similar names. Doh!

It's crazy how often I ask agents to use shadcn components and then they just... don't. They roll their own. It's much better when I have vercel and shadcn skills/tools/mcp installed plus stronger wording against rolling their own without stopping to confirm.

Working on finishing up my AI chat app. Going to have to change names, too many that are close to it. Working on final features for beta, fixing bugs and general testing, clean up, docs, etc. I should have a pretty good idea of release date after this weekend.