headcanon

This is a pretty cool setup! I've wanted to give my agents "a life of their own" as well and interact with it somehow, but struggle to find the balance between giving it freedom and being ok with the risk. How is this set up, do you just use something like a claude routine on a loop + context repo, or is it more sophisticated than taht?

I feel like I'm living in an alternate universe - my company has fully adopted agentic coding, we're working with:

- large codebases

- business critical software for clients that don't tolerate downtime

- agentic and classical features

We still collaborate well, we still care about quality, we still have good knowledge of the codebase, but I haven't written a single line of code myself in over a year.

We have not had a single incident that I'm aware of where we shipped a slop and it created downtime.

We're building a "software factory" that is currently handling low-level chore tickets and merging about 1/3rd of our PRs. A year from now I imagine it will be doing significantly more, and I don't think anyone is going to get fired because of it (if anything we'll keep hiring more).

I don't see this as an either/or thing. If there is some hidden ocean of slop that we're bound to fall into I don't see it anywhere. Where is the disparity?

That depends on the output's purpose: if the purpose is to produce readable text for a human that is _not_ me, like an article or document, then I care more about clarity, plain-speaking, and general register. If the goal is to accomplish a specific task, I don't care as much about the prose quality: I'll put up with "Honest Framings" and "load-bearing" since it seems to me that's the token that needs to be in the context for it to function.

I've been in the habit of pushing my claude-speak to codex to improve legibility, but only if I think someone is going to read it.

These artists are pretty cool:

- https://andreaciulu.com/ - https://www.rooperainisto.com/

You might feel differently, but I feel intention and effort being communicated here.

I will frequently post the output of a computer to slack as a part of an engineering discussion. For Claude output I'll treat it the same, wrap it in code blocks, and say something like "Claude's analysis". If its ok to post system logs, why is it less ok to post claude's output (also a computer), especially if I declare this? Is the difference that its plausibly human-level speech and thus breaks an implicit contract?

agreed, but IMO thats been true for most off-the-shelf skills I've seen. Instead I built my own "/meta" skill to improve my workflows, which I can use in targeted ways across my projects

I think AI optimism is an easier jump for those already immersed in computer culture, or at least the subset of that culture who view automation as a virtue.

For those either outside computer culture, or those within who might view "craft" as virtue, automating that craft is an affront.

Right now, the "automators" (pro-AI) folks seem to be pushing that culture onto the "crafters" and naturally getting pushback.

I see you haven't tried a modular synthesizer yet :) Getting back to the same "place" in a patch can sometimes be impossible, and it does feel "random" until you get the hang of it.

Two years ago (Dec 2023), I visited Japan and try to use ChatGPT (GPT-3 era) to translate Kanji signs, it was OK but not great. A little over a year later (late April 2025, o1 era IIRC), I visited again and saw a political rally with hand-painted-style Kanji signs. I took a picture and asked what the political party was about. The big "wow" moment was how it would write inline python code to crop analyze the image while it was thinking, then combine it with web searches to eventually find me the website with a summary of what they were about. It definitely hit home for me the potential of what these were capable of.

I would definitely agree that data center investments would need to be coupled with energy investments. If this could act as a catalyst for more (sustainable) energy production that would be a net win for all IMO.

Do you feel poison fountain is actually effective? To me it seems like free chaos engineering for ingestion platforms. Wouldn't it paradoxically harden data ingestion?

I've been trying to solve a lot of the issues brought up here, like personal automation, note taking, article summarizing/indexing, etc. What I've learned so far is that no level of AI-enabled automation will help me with my inherent ADD. All it means is that I have more started projects, but just as many finished ones. It has been a big enabler for things I'm already knowledgeable about, but the cognitive debt issue is real: A machine that thinks for me can't help me think better by itself.

Came here to say this as well. I've written and read full human-written essays on slack before AI.

With that said, I don't disagree with the article. Don't use more word when few work.

To use your analogy, I would say the "blanket ban" attitude would be more like wishing all viruses would just go away, or have never existed in the first place, which:

1) is an impossible and unproductive attitude, and

2) fails to recognize the important contribution to evolution, genetic diversity, and our immune systems that viruses introduced, not to mention the possible beneficial applications that could exist by understanding it.

Rejecting something without nuance makes you more vulnerable down the road because it prevents you from building an effective immunity. Engaging with it is the only productive way to mitigate its downsides and promote its benefits.

My wife bought a Neo and has been very happy with it. I was wary of the 8gb memory limit but she is running claude code doing web development with a reasonable number of tabs open and no noticeable lag, so I'd say its definitely getting a lot of mileage out of it.

It honestly seems good enough that it might cannibalize Macbook Air sales.

As others have said, the main benefit with Python over Rust is library support especially with ML features. The other gap as I see it with Rust is the lack of native flexible UI support. The nice thing about Rust though is it can serve as a very fast and stable core for an app and offload specifics to TS and Python as their strengths allow, so you get the best of all worlds.

My current goto for desktop apps is Tauri, which give us a rust backend and TS fronted (usually React). Local ML features can be easily loaded as a python sidecar. Production bundling can be a little challenging but it seems to work well so far.

Sidenote: Golang is also an amazing language for LLM use, I generally do most of my "infra" stuff in Golang over Rust, but either work fine most of the time.

same, I struggle to use more than half of my weekly, even if I max out my 5-hour windows regularly during the day.

I've been seeing much higher session limits late at night (US time). Workday usage struggles though.

I'm looking into how to structure my work to run some autonomous-safe jobs overnight to take advantage of it.

I've been trying to toe the line here myself, here's how I've been doing it. For context, I pay for a Max 5x subscription.

My main goal is to maximize my subscription token usage while trying to comply with the rules, but its not clear where the line is for automation so I feel like I need to be clever.

- regular development (most token use): all interactive claude mode, standard use case

- automated background development: experimenting with claude routines (first-class feature, on subscription)

- personal non-nanoclaw claude automations (claude -p): uses subscription token, but only called as needed (generally just fix something if something in my homelab infra goes does down, its set up to not fire on an exact cron time)

- other LLM based automations: usually openrouter API key, cheap models as needed

- nanoclaw: all API key based, but since its expensive I keep usage mostly minimal and try to defer anything heavyweight to one of the other automation strategies (nanoclaw mainly just connects my homelab infra with telegram)

I tried openclaw when it was released, but I preferred the minimalism of nanoclaw and replaced it. I have it on a mac mini now

For context: I have additional automation with scripts set up on the mini, some of them call LLMs to do things like summarizing today's news.

I have other automations that are agentic and just run "claude -p" (mainly just checks status of other jobs and fixes them automatically). Agentic automations are great because they can handle unexpected situations (at the cost of predictability). They're all sandboxed and we have control over tools for the most part. Any files it would write to are typically git-controlled so we have change records and rollback built in.

Nanoclaw acts as an agentic layer but combines it with the communication layer over telegram to make it interactive.

I use it to go through my centralized task list (currently beads in my main 'wiki' monorepo), give me nudges for todos, I can also send it pictures of say, food and it will fetch a recipe and sort it into the wiki via a general "inbox" skill (claude has it as well). Every day at 12:30 it will give me a mini "standup" of all my personal projects and todos, and once in a while will give me some thoughts based on my interests.

Its set up to do appropriate tasks with local models to keep token costs down, so far it doesn't seem to cost more than $10-20/mo, it would use less if I didn't drive it with sonnet.

I'm still experimenting with it, and trying to go slow, one thing at a time. I don't give it access to anything super sensitive yet, and try to keep it observable.

> Early career folk join some toxic startup and stay because the internet told them all CEOs are like this.

I literally did this 12 years ago based on this reasoning, its good you're trying to counter that with the next generation.

With that said, I do wish there was more discourse around systemic issues rather than the usual finger-pointing towards rival social groups. Unfortunately I feel like our language gets in the way, systems issues are more abstract, but "bad people" are more visceral and easy to talk about.

This is a big part of why I'm looking to develop a local LLM capability: having the hardware is a good start, but also developing the understanding on what the SoTA of local edge models can do, so we're not crippled if remote models stop being served, or at least some risk management.

It doesn't solve the problem of general LLM dependency (at the end of the day we gotta keep our brains sharp), but any LLM-based workflows aren't all of a sudden put at risk if we set up something that depends on it.

Kind of. I'm finding that my terminal window in VSCode went from being at the bottom 1/3rd of my screen to filling the whole screen a lot of the time, replacing the code editor window. If AI is writing all of your code for you based on your chat session, a lot of editing capabilities aren't needed as much. While I wouldn't want to get rid of it entirely, I'd say an AI-native IDE would deemphasize code editing in favor of higher-level controls.

I've been getting close to that myself, I've been using VSCode + Claude Code as my "control plane" for a bunch of projects but the current interface is getting unwieldly. I've tried superset + conductor and those have some improvements but are opinionated towards a specific set of workflows.

I do think there would be value in sharing your setup at some point if you get around to it, I think a lot of builders are in the same boat and we're all trying to figure out what the right interface for this is (or at least right for us personally).

Most knowledge worker use computers today to do their work, but we don't necessarily call them computer or software engineers. I think it will be something like that, but the economy will need to adapt and grow in order to accommodate it.

I agree with most of the other comments so far, the "magic" of the show is going to be hard to replicate.

My bull case for this is that Nathan Fillion and crew have had 20 years of exposure to this fact and likely know what they're getting into and how to do it right. The only question is if they'll be able to execute. I'm excited!

Thats exactly what we as software engineers do. We are constantly automating ourselves out of a job. The trick is that we never actually accomplish that, there will always be things for humans to do.

We're discovering so much latent demand for software, Jevon's paradox is in full effect and we're working more than ever with AI (at least I am).

If anyone wants a chuckle, I vibe-coded an endless supply of "synergizing paradigm" terms as a slideshow for a fake corporation. It's fun to put on in the background on a tv somewhere to see if anyone notices.

https://brightpath-global-solutions.com/

Edit: repo link: https://github.com/chronick/global-business-solutions

+1 for Tauri, I've been using it for my recent vibe-coded experimental apps. Making rust the "center of gravity" for the app lets me use the best of all worlds:

- declarative-ish UI in typescript with react

- rust backend for performance-sensitive operations

- I can run a python sidecar, bundled with the app, that lets me use python libraries if I need it

If I can and it makes sense to, I'll pull functionality into rust progressively, but this give me a ton of flexibility and lets me use the best parts of each language/platform.

Its fast too and doesn't use a ton of memory like electron apps do.

Yeah that would be nice. My thinking is that they don't want to cannibalize ipad sales.