harness initially annoyed me as well. i asked chatgpt why it’s called a harness and since then im now fully onboard heh.
the agent explained that the metaphor is about steering, a horse harness steers but it’s not technically the control, it’s a suggestion to the horse. The harness is about defining boundaries and such. You perhaps already know this, it’s just that in this instant-meme world, the fact that a word has a substantive meaning is comforting.
> The Chinese term for involution, "neijuan", began trending online in 2020 and was initially used by young people to describe the hypercompetitive and often self-defeating pursuit of traditional markers of success. Some of the contexts they used it in included questioning what was the point of working hard to get into a good school if the reward was working 996 hours (9 a.m. to 9 p.m., six days per week) in a tech company? If you were lucky enough to land a job, that is, in an era of high graduate unemployment.
> Though the term is far less commonly used in English, involution comes from a latin term which means "to roll or turn inwards". It was popularised by American cultural anthropologist Clifford Geertz in the 1960s - in relation to his studies of Javanese agriculture - to describe economic or cultural stagnation despite increasing complexity or effort.
> More recently, neijuan has become shorthand in China for the exhausting but also often futile and sometimes self-destructive grind of hyper-competition more broadly.The concept is now also linked to the country’s pivot from property-driven growth to an industrial complex encompassing a third of global manufacturing, which has seen more recources invested without any accompanying increase in returns. It's a race to the bottom.
I mean - it has a landing page advertising its features. It has an enterprise tier and a "sales team". It has a merch store. It's clearly not meant to be seen as "just for fun."
Wouldn't an agent in any capacity need a sandbox runtime to do agentic work? When you say "sandboxes are too slow" do you mean needing to boot the sandbox to get back Agent state? Is the idea that state/memory/context and runtime env are separate concepts in rivet?
(I'm a professional developer but don't do any orchestration so trying to understand from basic mental model pov)
There is a notion right now of decoupling the agent loop from the sandbox, so the sandbox isn't waiting for request-response from the agent. This would probably be 90+% of all the agent work.
If you combine this with products like Sprites from Fly.io (or OSS version - agent substrate), where you pay only per execution request in a sandbox, you can have huge savings compare to the event loop inside the sanbdox. You will be billed only for request executions (tens of milliseconds per read/write, hundreds of milliseconds per search, etc.) and zero for all the waiting time during the the inference phase.
Also, if you don't require a real Linux sandbox — you might actually emulate the environment, e.g. just-bash idea with linux-like environment to give agents something that looks like their native habitat, but under the hood, it can be just CF Durable Objects SQLite or whatever else.
Probably(?) almost all knowledge work doesn't require a Linux sandbox. You can just build a document reading and editing agent that does tool calls and connects to other products via MCPs, APIs, etc. And lightweight actors make a lot of sense to me in these cases.
> Wouldn't an agent in any capacity need a sandbox runtime to do agentic work?
Developers on Rivet use 3 different patterns:
- Just agents: many agents just need tools, such as customer support or personal agents
- Agents with lightweight sandbox: Rivet provides a V8 / WebAssembly sandboxing library for running JS & Python with a filesystem, which covers most workloads like data analysis, file processing, evals, and code mode (https://rivet.dev/agentos/)
- Agent outside of the sandbox: expose the sandbox to the agent as tools, which is better for blast radius, trust boundary, and orchestration (see below)
> Is the idea that state/memory/context and runtime env are separate concepts in rivet?
Exactly, Rivet Actors support running the harness outside of the sandbox. It's often referred to "separation of head from the hands."
> When you say "sandboxes are too slow" do you mean needing to boot the sandbox to get back Agent state?
Specifically coldstarts. As LLMs are getting faster and have more realtime use cases, sandboxes are become the bottleneck on time to interactive for an agent.
im so curious what specifically makes this possible. I have no experience with conductor but your comment reminds me of how impressed I was with the development speed of parse.com. I'm likely dating myself; it was the first serverless/firebase backend concept that supported mobile app development. I wasn't even a user I just noted how prolific the feature rollouts were. So polished too. They got acqui-hired by Meta and I'm sure the founders did well.
Anyway, for anyone reading, just a not developer-led products. How do they do it?!
disagree. the web specifically is built as a dynamic medium for consumption. The fluid nature is the point. Yes, it is harder for developers to force a consistent experience, but as a professional web dev, we should not treat the web as a fixed medium.
Javascript is notorious for not having a standardized way to do things. It doesn’t have a strong opinion on anything really. So I buy the argument that frameworks displace or even increase ways to do things. React patterns, SPA, nextjs, its all just too much.
Point is frameworks or not, Agents make a mess in javascript, so maybe it’s actually an improvement to establish the standard lib as the foundation and steer Agents from there. Adding frameworks expliodes the surface are of options.
Python is worse about this proliferation, despite the advice of its built-in koan (see: `import this`).
Sure with JavaScript's library a-la-carte fashion there's a proliferation of idioms far beyond eg Ruby's hegemonic "convention over configuration", but nevertheless React is still more of a lingua franca than Flask/FastAPI/SQLAlchemy ever was or will be. The most popular Python libraries do a thing and otherwise proscribe nothing.
JavaScript idioms pare down the universe of possibilities in a way that some programmers find claustrophobic, and Python is the wild west!
I know it's popular to characterize JavaScript as the wild west, but in that case it's a proliferation of abstractions, whereas with python, it's a proliferation of implementations, ie primitive obsession.
But I'm really putting far too fine a point on these differences. In the grand scheme of things, they are nigh indistinguishable scripting languages.
I'm dating myself a bit here, but I'm reminded of Douglas Crockford's book, JavaScript: The Good Parts. The idea being that under all of the cruft and foot-guns, JavaScript actually had all the elements of a good language. You just had to be disciplined about using only a subset of the language.
Isn't that essentially what we get with frameworks? I'm still unconvinced that frameworks are a problem with agents. If anything, pointing an agent at a single framework reduces its API surface, compared with the standard lib and every permutation of possible solutions that comes with it.
The combination of ES5 (and successors), Typescript, and community consensus resulted in everybody using “the good parts” without even having to think about it. Frameworks, in proportion, didn’t have to do much
I do agree, it’s just likely really hard in practice to keep agents disciplined. The most popular frameworks are quite substantial. For example my experience with agentic react development was both amazing - in that the sheer volume of drop-in interfaces allows you to create stunning interfaces previously reserved for startup darling design teams - and nerve wracking because the way agents integrate everything is just layers upon layers of hooks. hooks provide intentional loosely coupled plugin interfaces, but it’s not sane at all to take these to the extreme with literally 8 layers of state. im convinced it’s unsustainable and ridiculous. These AI apps are basically un-designed internally. Can’t understand how they’re maintained.
The element of a good language is that it doesn’t have the foot guns. If I shit on your birthday cake, you don’t eat it because a reknowned baker wrote that, “underneath the feces, it had all the elements of a good cake”.
I'll definitely be taking a look, but it seems that the videos are still hosted on Youtube?
The main issue with Youtube are the ad rolls. The video content itself can be great. But I wouldn't wish those ads on my worst enemy. It's mind poison. Straight poison into your consciousness. AI generated ads for every conceivable image and status vulnerability. And they now do this weird thing where the ad is delivered as a pop song. Like a Taylor Swift engineered pop song for a Korean face mask. The song is 10 mins long about how this woman was happily married then divorced then saw her husband with a woman half his age and how that made her feel and how her self worth was tied to her marriage and her looks but also she's not just that but also somehow the solution is a Korean face mask. It was really traumatizing for me. You can see how much it infected my mind. It's poison.
Thank you for giving me a glimpse. Reading half through your comment made me realize I never saw YouTube ads since the AI cancer, naturally, I wondered what it feels like.
I watched the entire thing to provide meta-commentary as I just did. I explicitly try not to solidify my position with knee-jerk reactions to the world. So yes I did watch the entire thing to fully qualify the poison-experience.
What you can do is start a channel where you feature your emotional reactions to each long-form ad as it rolls on the screen. Then, you can purchase the items from the ads and explain how they totally were not the solution they promised to us. Then, having monetized your reaction/review channel, you should have enough to pay for YouTube Premium every month!
nerd-sniped me on this one. mainstream cool isn’t a thing. i think you mean successful. mainstream values success because it’s the vanilla proxy for a vanilla form of “value”.
so cool has always been about the “iykyk” kinda value. once too many people know there’s nothing to know anymore.
I think your cool only applies when you get quite granular. OTOH, everyone knows "teenager slang" is cool and "working in tech" is (now) cool, or "being a brain surgeon" is cool. So as I said in my original comment, those are two different "cool"s entirely.
I've been happy with fastmail. Their product and UX looks a little dated even - in a good way I realize. It's not trying to be anything its not.
Also iCloud paid accounts allow custom domains for email. The Apple mail product is way worse but it was amazingly simple to set up and I am happy I have my own domain.
Apple is extensively and heavily using GenAI internally, and I've not got confidence in their product quality going forward (macOS is still declining in overall quality, from my experience)
the agent explained that the metaphor is about steering, a horse harness steers but it’s not technically the control, it’s a suggestion to the horse. The harness is about defining boundaries and such. You perhaps already know this, it’s just that in this instant-meme world, the fact that a word has a substantive meaning is comforting.
reply