Been running it on a locked down Hetzner server + using Tailscale to interact with it and it's been surprisingly useful even just defaulting to Gemini 3 Flash.
It feels like the general shape of things to come - if agents can code then why can't they make their own harness for the very specific environments they end up in (whether it's a business, or a super personalized agent for a user, etc). How to make it not a security nightmare is probably the biggest open question and why I assume Anthropic/others haven't gone full bore into it.
This is admittedly dated but even back in December 2023 GPT-4 with it's Vision preview was able to very reliably do structured extraction, and I'd imagine Gemini 3 Flash is much better than back then.
It reuses previously generated code, so tools it creates persists from session to session. It also lets the LLM avoid actually “seeing” the tokens in some cases since it can pipe directly between tools/write to disk instead of getting returned into the LLMs context window.
I've been using it daily, so far it's built CLIs for hackernews, BBC news, weather, a todo manager, fetching/parsing webpages etc. I asked it to make a daily briefing one that just composes some of them. So the first thing it runs when I message it in the morning is the daily briefing which gives me a summary of top tech news/non-tech news, the weather, my open tasks between work/personal. I can ask for follow ups like "summarize the top 5 stories on HN" and it can fetch the content and show it to me in full or give me a bullet list of the key points.
Right now I'm thinking through how to make it more "proactive" even if it's just a cron that wakes it up, so it can do things like query my emails/calendar on an ongoing basis + send me alerts/messages I can respond to instead of me always having to message it first.
I went down (continue to do down) this rabbit hole and agree with the author.
I tried a few different ideas and the most stable/useful so far has been giving the agent a single run_bash tool, explicitly prompting it to create and improve composable CLIs, and injecting knowledge about these CLIs back into it's system prompt (similar to have agent skills work).
This leads to really cool pattens like:
1. User asks for something
2. Agent can't do it, so it creates a CLI
3. Next time it's aware of the CLI and uses it. If the user asks for something it can't do it either improves the CLI it made, or creates a new CLI.
4. Each interaction results in updated/improved toolkits for the things you ask it for.
You as the user can use all these CLIs as well which ends up an interesting side-channel way of interacting with the agent (you add a todo using the same CLI as what it uses for example).
It's also incredibly flexible, yesterday I made a "coding agent" by having it create tools to inspect/analyze/edit a codebase and it could go off and do most things a coding agent can.
100% - sharing CLIs with the agent has felt like another channel to interact with them once I’ve done it enough, like a task manager the agent and I can both use using the same interface
Cool to see lots of people independently come to "CLIs are all you need". I'm still not sure if it's a short-term bandaid because agents are so good at terminal use or if it's part of a longer term trend but it's definitely felt much more seamless to me then MCPs.
Pydantic-AI is lovely - I've been working on a forever, fun project to build a coding agent CLI for a year plus now. IMO it does make constructing any given agent very easy, though the lower level APIs are a little painful to use but they seem to be aware of that.
Part of the reason I switched to it initially wasn't so much it's niceties versus just being disgusted at how poor the documentation/experience of using LiteLLM was and I thought the folks who make Pydantic would do a better job of the "universal" interface.
The heat gun mosquito things that some tech folks were mentioning on Twitter. Always get quarter sized terribly itchy bumps for each mosquito bites and using it makes them essentially itch-free immediately.
Under $1000:
Weekly house cleaning. Such reduced cognitive load/increased free time to not have to clean all the time, think about cleaning, etc especially with a toddler.
My personal (overly biased view after reading Chip War recently) take is pretty much there, seems like a lot of the same early dynamics of semiconductors are playing out here.
Very large R&D expenditures for the next iterations of the models at the leading edge (the "fabs" of the world), everything downstream getting much cheaper and better with demand increasing as a result.
Like a world where Claude Opus 3.5 is incredibly expensive to train and run, but also results in a Claude Haiku that's on net better than the Opus of the prior generation, occurring every cycle.
One fun detail - I don't know any Typescript nor have I ever coded a VSCode extension before, so this was all made going back and forth with Sonnet 3.5 + digging through the VSCode repo to figure out how their new proposed shell integration API works.
Virtual actors are so nice, Dapr[1] uses them as well. You can get hack together something like a virtual actor with Ray too but I'm hoping their actors eventually evolve into something like virtual actors eventually.
You can do some really cool things now with these models, like ask them to extract not just the text but figures/graphs as nodes/edges and it works very well. Back when GPT-4 with vision came out I tried this with a simple prompt + dumping in a pydantic schema of what I wanted and it was spot on, pretty much this (before json mode was a supported):
You are an expert in PDFs. You are helping a user extract text from a PDF.
Extract the text from the image as a structured json output.
Extract the data using the following schema:
{Page.model_json_schema()}
Example:
{{
"title": "Title",
"page_number": 1,
"sections": [
...
],
"figures": [
...
]
}}
https://github.com/caesarnine/binsmith
Been running it on a locked down Hetzner server + using Tailscale to interact with it and it's been surprisingly useful even just defaulting to Gemini 3 Flash.
It feels like the general shape of things to come - if agents can code then why can't they make their own harness for the very specific environments they end up in (whether it's a business, or a super personalized agent for a user, etc). How to make it not a security nightmare is probably the biggest open question and why I assume Anthropic/others haven't gone full bore into it.