AI loves smoking guns

Pop quiz: was this post 5 days ago? or... 15 years and 5 days ago? Man, how time (I mean 5 days) flies.... cc @frommww @treeder @dbasch

Recent improvements to my claude dashboard: - Added a governance dashboard to track auto-loaded context files, permissions, and hooks. - Added support for cmux and ghostty terminals I still use this everyday to easily track and resume sessions. Not a huge fan of how cmux or even the claude agent view does it yet. github.com/carimura/claude-d…

One of the biggest challenges I'm seeing with agents is how confident they are even when wrong. Somehow my agent decided that Cloudflare containers had a max concurrency of 10 and wrote all this pooling logic to accommodate, resulting in queue backlogs and an infinite loop of trying to tweak the pooling logic. Once I got involved, read the docs with my own eyes (the horror!) and challenged that assumption everything got simpler and smoother and I watched 100 containers kick up to process the backlogs. I then reminisced of the IronWorker days when customers would spin up thousands of workers to chew through queues and get to that “holy s**t” moment. Ahhh the good ole’ days!

small tip: you can set claude code to "ANSI colors only" theme which will match your terminal theme. Just choose light or dark and it will still honor your system settings (probably a product of vibe coding). Then in ghostty you can clone the theme and tweak all of it to your liking. Combine this with a theme-switcher I created [1], and boom. [1] github.com/carimura/theme-sw…

Get ready for a strange world where products aren’t designed for your eyes, but for agents. An app I'm building starts by putting objects in a Cloudflare r2 bucket, but my agent harness wasn’t looking at the whole wrangler response and threw out some really important context: “Resource location: local”. So then it spun off in a complete tangent writing a bunch of confused code and looking for known issues because it couldn't find the file in r2 when it was really just sitting there in a local r2 cache. To be fair, this is likely a harness issue because the wrangler CLI functionality changed between versions but the harness was “looking” (aka grepping) for the old format. Point being: agents don’t “see” the same as humans and (surprise) we shouldn’t assume they are right.

We should be nervous about how much power the AI models have in discovery. I’m doing some car speed tracking using a camera on our property and Opus chose an app stack (Python:3.12-slim base image, UV, FastAPI, uvicorn, ) ML stack (Yolo11s, Bytetrack, ONNX) cloud architecture (Cloudflare R2->queue->worker->container->D1). In the before times, I would have spent a whole day researching these pieces, testing them, talking to friends, and building a plan. Now? I gave it a nudge about the cloud arch because I had an opinion, but otherwise, turned “auto mode on” and let it rip. Took a few hours to build. I never opened a browser let alone a search engine. This will start to apply to every field: education, health, politics, history...

In the serverless days with IronWorker, IronMQ, and IronCache, our vision was to build apps without thinking about infra (inspired by App Engine). But then something happened. 100 JS frameworks, 400 cloud services, and Kubernetes all came around, it seemed more complicated than ever. Now with AI it feels like some of that original vision could become reality because it can choose an AI-native platform and master all the APIs. An alternate theory is that advancements in infra just push the boundary for what CAN be built further. There are only so many TODO list apps needed in the world.

9am - Chad builds a dashboard to manage code sessions. github.com/carimura/claude-d… 2pm - Anthropic announces agent view. x.com/claudeai/status/205394… This is not the first time this has happened. What a machine they’ve got going over there!

Coding before: sit down, open project, look at code, think for a bit, write some code, feel satisfied after some moderately-predictable number of hours, go get lunch. Coding now: sit down, open project in claude code, decide you need a dashboard for claude code, build one, but your terminal could be better, build a new one using libghostty, ponder token spend, build a dashboard to track that, go back to original project you wanted to work on but decide you need to eat, so not really feeling satisfied, go get lunch.

The process of software reminds me of a Bob Ross painting. You can’t “just paint the tree” too early. I haven’t been able to jump from a claude code terminal to say, lots of hermes agents reading tickets. There’s too many paint splotches to get on the canvas first. I’m guessing this is different once the software matures a bit? youtube.com/watch?v=lLWEXRAn…

All the systems downstream from agents pushing code are hurting. GitHub is down every day. Railway's "deploy queues" were so far behind they shut down deploys for low-cost accounts. Wild times.