Honestly the fact that we're still modeling agent harnesses like chat and strapping them into shell sessions instead of building async actor systems with sandboxed OS functionality access actors is bananas to me. So much easy wins can be had just by building on right abstractions... Hopefully will get enough time to play with this idea soon on my own.
https://tau-agent.dev/ in my setup. Many sessions each multiagent, in different sandboxes, sending messages. Also nothing in the main article that Tau wouldn't have. I just don't have stamina for more "marketing". Async tools ... sooo basic. :D
I'm actually taking the time to hash out the details to do this as a side experiment project, but basically what I read this project as showing is that if you leverage asynchrony you can let agents be more efficient - and you get this "for free" if you model your "agents" as actors that interact with the system through message passing. Then all the system operations become messages to different actors, need to read a file => message the fs actor => get reply from FS actor as a message when it's done. You'd probably need some out of mailbox ways to share blob resources and sockets for realtime (audio), but for the most part simple message passing should handle most of what LLM agents do.
Actors can have identities and roles for RBAC, etc. you - cross agent communication is the same as sending any other message to a actors.
Not to mention that actors can be on your device, another device, etc. whatever the router can resolve - it's transparent to the agents.
I think you nailed it. I have my own agent harness that's designed with inspiration from erlang/elixir. It seems to be a natural fit for indeterministic output and failures.
I’d think the approach in this repo gets close to that, no? I also think LLM harnesses should just support typical async i/o.
Let a model get notifications when something it has accessed/ subscribed to changes, a tool produces output or it gets some other kind of message (for example by another AI or flesh agent). Wait until the next turn or wake it up. It can still decide to do nothing and wait.
That's the reason I'm excited about the orchestrator because it would let me build my ideal UX on top.
If you move away from the idea that agents are chat streams and treat them as processes/actors then you can start letting agents represent themselves/build their own interfaces.
And if you build enough introspection into the protocol because everything is message based you can have other agents build interfaces for them.
So like either standard GUI, or a voice assistant talking to you and delegating to agents, etc.
It's not an idea it's a way of thinking about agent systems - basically an agent OS.
Copyright is artificial scarcity rationalized by arguing that producing novel intellectual work is valuable, but requires substantial effort that can't be recouped, so we have to incentivize it somehow.
LLMs and AI are changing that proposition substantially - human effort involved in producing copyrightable content is getting reduced constantly to the point that if we abolish copyright entirely we'll still have more content than we could ever hope for.
AI/robotics eliminating scarcity of physical goods sounds very far fetched but in the intellectual space it looks very very plausible in the near future - so it could be time to abolish IP laws soon, especially if AI manages to advance enough in R&D and research space.
If there is no legal guarantee that human creativity can pay off we are starving art and humanity from its inception. If ordinary people cannot participate in the act of creation, you get exactly what hollywood has become.
Sorry, but this reads like a mouthpiece exactly from those companies that benefit the most from having no copyright and I doubt your have thought this actually through.
Lack of way to capture value from intellectual labor is considered a market failure that leads to suboptimal market results for consumers, but with AI that argument becomes very weak.
Sorry but the point isn't to create artificial scarcity just so intellectual labor is well off, that's a negative side for the consumer that was considered necessary tradeoff. Market economy should be about providing the most value to the consumer.
Disclaimer - I was never a fan of IP laws despite them working in my favor, with AI I can see them finally being abolished.
We could imagine, as an extreme case, a technologically highly advanced society, containing many complex structures, some of them far more intricate and intelligent than anything that exists on the planet today – a society which nevertheless lacks any type of being that is conscious or whose welfare has moral significance. In a sense, this would be an uninhabited society. It would be a society of economic miracles and technological awesomeness, with nobody there to benefit. A Disneyland with no children.
There's also a recency bias, where you ignore how valuable some decision was in the past and ignore the fact that you maybe wouldn't even get to the place you are if "you made a different decision in the past".
I can't count how many times I've said "this is really bad, I it is worth rewriting to fix all the issues", only to discover that there were good reasons for all the past decisions and so we end up with the same mess as before - except that now I know what it must be that way.
Not always, but very often people in the past had good reason for what they did.
Those are the comments that should persist in the code. I hate when the AI edits and removes my "why" comments. I want the refactor to make it less messy but keep why it's that kind of mess.
I just put a comment in yesterday for this very reason - there was a flag I had set to false, and then looking at the library docs for something else thought maybe it should be true but that doesn't work how I would have implemented it if I had created the library and written the docs how I did.
It is a very capable library, and when I first started using it the docs were more Oracle docs looking - but I could find what I wanted easily. Now, it is much more annoying to delve through but looks more modern.
But of a tangent but I think cognitive skills are starting to become like physical skills. If we don’t move our bodies, we waste away physically. If we don’t do hard cognitive work sometimes - like writing and debugging code - I worry our minds will atrophy.
I don’t have a problem with cars. But walking is still good for us.
Try, but what skills are work keeping? We need something cogntive, but not everything. I know a few people who blacksmith as a hobby (often for the physical exercise as much as the work), but most people are happy not knowing how to do that job. I know how to set the air-fuel ratio on a gas engine, but I'm glad I don't need to tweak those parameters while driving (unlike a 1910s car where you did), and I won't miss oil changes on my cars as I move to electric.
If I'd accidentally written a comment asking whether the intellectual skills I am losing are worth keeping, and make 2x spelling errors in a single short sentence, I'd find the unintended irony hilarious.
(but that's me, maybe you don't find accidental errors that you make to sometimes be funny)
Meh I just replaced a battery in my 14 pro (and the charging port) for ~200 EUR. I didn't want to upgrade yet and I would not even be considering upgrading if the device had USB-C
We heard this since Uber days and all it did was break the old taxi systems and brought taxi service to a wider audience. As a consumer I benefited immensely from Uber.
Likewise for AirBnB and hotels.
Not saying that these don't have secondary effects on the society - but as a consumer they were and arguably still are pretty amazing.
Me and quite a few others are getting real tired of the constant name-dropping of mega-corporations that do not care about your safety, health, or well-being. This does not resonate with anyone except a small circle of wealthy tech elitists.
If they can make a profit from those things, great, but they will also not hesitate to make a profit from your suffering as well. Ask the lung doctors of the people living next to xAI's data centers if their patients are better off than before.
You benefitted temporarily. Those offering the service are often worse off. Various regulatory bodies had to beat these companies into submission, that no, you cannot hire all your employees as contract workers and deprive them of their benefits. After the competition is murdered, they will raise their prices over and over again.
All these companies thrive on the externalization of costs and killing an economy to then use their monopoly position to maximize money extraction.
I think in this case, the contract workers are the ones taking the L, generally. Taxis are not overpriced just from a lack of competition, but because they are an actual full time job with all that comes along with it, and not well.. part time gamified bullshit where you fight for scraps.
Most of the does EU has Uber now, and Wolt and some others, so it's not like we avoided anything there in practice though.
You can easily setup codex rules to auto approve local stuff but gate external effects like push/jira write/curl - works better for me than full yolo mode - depends if it's a solo project or working with a team.
I'm sorry but no - output quality matters a lot more for me.
Just yesterday I tried to use Google antigravity to do a side project I've had on the back burner for 10 years now. Gemini flash is insanely fast - at first I was amazed at how quickly I was getting responses, and it seemed to hold it's own in technical discussion, although sycophancy is next level. But then when I actually let it do the coding part it was just drivel. I wouldn't even bother improving that code - like cleaning up after a lazy unskilled coworker - throw everything away and start over because the foundation is just leading in bad direction.
I spun up Astra on the same problem and although it was sluggish in comparison, and much more pedantic about irrelevant details - the feedback/pushback was actually meaningful. The implementation PoC also took tweaking but we got on the same page really fast.
Gemini Flash 3.8 was just producing garbage ultra fast, Astra could actually be steered into a direction I want and it provides valuable/insightful feedback.
I don't have infinite reading capacity/mental stamina - I would rather the model take it's time and let me see something high quality rather than get bombarded with garbage. If it can be faster that's great - but I'll always default to smarter model. The only exception is stupid trivial tasks like log analysis and similar.
Except there's a huge gulf of self-hosting and using API hosts - no way you can reach the economics of a shared host. Privacy is a problem but you can chose who you host with and where it's hosted (which jurisdiction).
When privacy/compliance really starts to matter it's up to the client/business to provide you with tooling - you're not running that on your own hardware anyway.
So the local AI for individuals is just a hobby/gimmick at this point not a rational decision. Self-hosting for business is a different story.
I'm not sure. The problem with the cloud llm's is they are complete black boxes that change frequently and randomly day by day.
If you run Qwen 3.8 on your own hardware, every single day, it's the exact same model running in the exact same way.
Yes, it's nowhere near as "smart" as the cloud based models. But it's consistent.
So the workflows and "ways of working" you create will work mostly similar day to day.
With Claude/OpenAI you frequently find days where the models are useless, and days when they are out of this world.
So I guess the choice comes down to:
1. Randomly the smartest thing on the planet with unpredictable rate limits that is mostly amazing, but frequently messes with your workflows
2. A really good local coding model that is consistent every day with no rate limits
I'm not sure. My gut feeling is maybe the right answer is a mix of both.
Gambling on the biggest models, hoping they are working smart that day, when planning or doing very complex work. Then doing most of the tasks/daily work using local models??
You can run any open model on a shared API host via OpenRouter and pin to which host you want to go for the quant/privacy/etc. mix you care about. You can pay them directly if you don't want the OpenRouter overhead - but the convenience of switching, having one invoice, etc. is worth it IMO
It's not closed hosted models vs open local models, it's hosted open models vs local open models where the math doesn't work for local LLMs.
The only local inference use-case I can think of is porn generation (because most providers don't want to deal with it) and illegal shit like hacking to minimize the tracing.
And if you're super paranoid - but honestly giving sensitive info to LLMs in any scenario is a gamble.
If you game and can use your GPU I guess then it works as well but models that fit into a gaming GPU suck too much to bother IMO.
reply