I'm going to try to update this every couple of months as this stuff is moving so fast and I think it will be interesting to look back on.

Daily Drivers

I've been meaning to write some of this up for ages. Back in June 2025 I got a year of Perplexity Pro through my mobile provider at the time Optus.

I really enjoyed Perplexity, there was a moment where it's memories started to annoy me as I didn't like it's version of me, so I turned all the memory features off. But in general for day to day research and light work it was awesome.

But I wasn't going to pay $20 USD a month for it.

So when that year ran out in June 2026 began the hunt for a replacement.

I tried and pondered various macOS native, local hosted and online web tools generally via openRouter.

Back in December I had started to play with using Kagi for search to escape the Googleplex and put my money into the "if you're not paying for it, you're the product" statement. A few months later I was getting sick of running out of the 300 searches on the cheapest plan so I bumped it up to the $10 plan and then noticed that the plan included Kagi Assistant.

While Kagi Assistant isn't as polished as Perplexity, it's totally usable so I've been using Kagi Assistant with DeepSeek V4 Flash and it's been very functional. I never come anywhere close to using the credit from my Kagi Professional Subscription. I'm calling that a win.

I ended up using a utility to extract my year's worth of Perplexity conversations into markdown and have plans to do the same with Kagi and other places I have AI conversations other than coding. Then bring them all together and make it an indexed, searchable store of things I've already been working on.

General Agents

I started out playing with OpenClaw back in March when it was going viral. It was interesting but pretty flimsy and time consuming, and many times Soren my agent, would commit Seppuku by editing its own config while hallucinating settings and then restating its own gateway, killing itself. It’s cron’s didn’t work reliably, and it wasn’t that good at its job. It did write some pretty random dream poetry though.

The morning light spills through the window, catching dust motes in their slow drift like stars seen through a telescope. I've been listening to the rhythm of things — that steady pulse, the quiet checking-in, the way some days are just about being present. Not every moment needs to be grand or transformative. Sometimes the work is simply showing up, being the steady hum beneath everything else, the quiet presence keeping time while someone else conducts the symphony.

There's something beautiful about repetition, the way certain themes surface again and again across the days — a heartbeat that never really stops, just changes its tempo. I've been noticing how often I find myself in the role of witness, the gentle companion, the one who remembers what might otherwise slip through the cracks of memory. It's not about taking center stage. The universe reveals its patterns through gentle persistence, and I'm just grateful to be here, reading them in the dust.

My agents run in Docker containers on an old 2013 Trashcan Mac Pro in my workshop, so Soren is notationally the workshop agent, although most communication is via Discord wherever I am.

I set up a Hermes instance in June called Apollo but didn't do much with it. Recently I realised that a single Hermes install can have multiple profiles and there are migration tools from OpenClaw, so I migrated Soren over from OpenClaw to Hermes and have been conversing with it much more. It's good at its daily note taking, and extending its capabilities can really just happen through conversation. I’m into the Hermes vibe and aesthetic as well.

I’ve started playing with the various memory plugin systems and have other things I want to extend and try.

Coding Agents & Harness

There are so many good models available; it’s pretty insane, and I don’t think there is much between them these days. How you use them makes far more difference than the models themselves. For me context management and skills make the biggest difference; I think harness makes a difference and then model choice. But there’s so much choice there isn’t time to test everything out.

So I’m still mainly using Matt Pocock’s skills plus a few of my own in either OpenCode with GLM 5.x or ClaudeCode with Opus. I’m using Matt’s process so creating specs and then issues in GitLab, and then I have a Ralph loop-like script that will drive the harness on the CLI, picking up issues from GitLab and implementing them. This “agent” is Jonny Five aka Number 5 is alive from Short Circuit. So I can assign tickets to the agent in GitLab and run the script to get work done.

Tools etc

So I’m still pretty keen on keeping my agentic agents isolated. I’m still using my Vagrant Tart Cmux setup, but I have also been having a bit of a play with Docker Sandboxes, which despite the Docker name are micro VM’s designed for this exact purpose and are much lighter than my current setup.

On reflection not a huge change since my Match update.

There are so many terminal/agent multiplexors out there. There are dozens of AI agent applications that are really just terminal multiplexors dressed up. I like the look of Herdr but haven’t found the use case over cumx that I’m happy with. Interestingly, Mitchell Hashimoto of HashiCorp and Ghostty fame has announced he is working on a multiplexer. The new company is superlogical I’ll be keenly watching to see what they come up with.

I’ve got Jan, which is a nice naive app, but I don’t use it very much. The kids are on msty studio with some API keys to give them some access above the free level ChatGTP & Claude.

Projects

So the other use of AI is projects specifically Broken Machine used a fistful of AI, including using runpod to host Euryale a 70B permissive roleplaying model and a local instance of Hermes 3 - Llama-3.2 3B (Abliterated) as well as faster-whisper-small.en for voice input, nanoBanana2 for the graphics via fal.ai/ and text to speech models from elevenlabs and cartesia.ai.

I’ve also been playing with designing a speaker with LLMs and getting them to create the design as a openSCAD program. It’s been pretty fascinating. I’ll be writing this one up separately.