Pi Is the Last Harness | Hacker News Reader

Pi Is the Last Harness

destiner.io

32 pointsDestiner38 comments

I don't think we've seen the last of anything in 2026/27. Lots of interesting projects are yet to start. I'd bet a good amount of money that some of the next Google/Microsoft-sized businesses will come from ideas in the AI harness/tooling space that no one has even noticed yet. No one is building them today, well, not exactly anyways.

And no one will see it coming except that one person in some remote part of the world (I don't say the US because, post-LLMs, I hardly think there's anything unique to US tech anymore. The best part was the sheer technical talent in the US, but that's worth a lot less with modern models and systems).

So no, I don't think we're anywhere near the end of innovation just yet. These will be some fun times, and I don't think the current way of building stuff will last too long. We'll definitely have better tools and ways of doing stuff, or so I hope. Keyboards have honestly lasted a lot longer than I would've expected, even though we've had STT tech for decades now.

Even today, office memes aside, it's just easier to type things out instead of speaking. For most people, typing is faster or fast enough anyway. Except when I'm on my phone.

Agree it feels super early for a lot of these tools. Don't think we have even scratched the surface of what working with LLMs will look like in 10 years. Just think about how many user interfaces can be completely replaced.

> Just think about how many user interfaces can be completely replaced.

Which ones? I don't see a single interface that can be replaced reliably. Complimentary maybe, but definitely not replaced. Would be interested to hear your opinion. It seems that even coding harnesses are a poor replacement for an ide. Drastically less information and if you think code doesn't even need to be read to be understood... Well, I guess we will just have to agree to disagree on that one. I think it needs to be engaged with at least as much as it was when written. Reading is the bare minimum.

A large part of software and websites I used before have been replaced by talking to LLMs. Just today I did some video editing, I would previously used OBS studio, but now I just prompt Claude with what I want, give me 10 options to pick from. Then my IDE, I don't use it anymore. Diagramming tools? Nope, prompting works better. Image editing? Prompts. Web design tools? Nope, prompts and give me 20 live previews to iterate on and narrow down. Web consoles to manage my servers? Nope. Excel spreadsheet for workout planning? LLM does calculations now based on my tracker API. I could go on.

So yeah I think we disagree. It feels so much better to me not having to switch between different tools anymore. When I read code it's not in my IDE but PRs on GitHub where I leave feedback for the LLM. I just tell it to keep iterating and monitoring PR comments until I approve.

As much as I would like to say I agree, I don't really like the LLM ui replacements. I absolutely hate them, they are just horrible, claude code itself is a horrible way of producing code/output you aren't seeing the logical pieces of your project you are largely stuck in a text box with a diff viewer and unable to parse what's happening.

Something like codex desktop is slightly better but still a horrible interface for developing software in any form of serious capacity.

Sure it works because the friction of manually working is just so damn high. And there aren't good alternatives, the jump from pi to claude to codex being almost negligible for most folks is a very high signal this entire pattern just doesn't really work.

I don't yet know where we will land or will we suffer with the current bad type text in a box ux for the next few years but I am certain this is not it.

I am still trying to work out what is it though, so well we will see when we get there.

Not using tools is a good example of the issue, you give up a lot of control to do that, sometimes it's good enough often it's not, and often it's worse because you don't even realize how much control you have given up.

I don't think we will do natural language interactions with our computers, I honestly feel like we will do something more intent based. If LLMs can know from our words what we want to do, then they can certain infer from what we are doing stuff we are about to do.

Think auto complete but for user actions, without taking away user control and understanding intent almost perfectly. Maybe it would require neuralink or something to get good enough, but until that happens I don't think we are moving away entirely from our UIs/UXs entirely.

Some people will, the low skill, low requirements, but no-code has existed for a very long time for the same reason, visual scripting is a similar example.

A video made by LLMs by using Premiere Pro is always going to be worse than what a professional makes entirely on their own in Premiere Pro not in terms of quality but in terms of control hence intent.

Similar issues with software, you can not perfectly describe what you want to another person perfectly that just doesn't happen, otherwise we wouldn't have had companies with more managers than engineers it seems like.

There is a lot of loss of intent, imagine if every action was something where you lost some of your intent, then after 100 requests does any of your intent even remains in it.?

This might be too hypothetical but I don't see LLMs replacing UI in the short term, but I don't know what happens in longer term.

But I do think LLMs will change how the software the interfaces and tooling works almost entirely.

Today it's much easier to llm rather than sql. It's not even a question of whether it's 10% better or something the cognitive load of writing SQL is high enough in data analysis that LLMs are the only solution.

LLMs enable some pretty interesting ways to query data itself, you can ask the data to query itself, imagine being able to prompt inject the data.

But on a more useful case imagine being able to let data build itself, when ever an incident happens, LLMs can enable flags to collect more data automatically based on incidents add reporting apis and telemetry and so much more.

Incident analysis should have changed but it hasn't yet, a human for some random reason still needs to ask an LLM to do the work, then understand it somewhat and then reply to others.

It really does come down to control and intent. Some people want nothing to do with either. To those people, creation is not an act of service or creativity, merely a stepping stone on their road to mediocrity.

As as society we have decided this is okay, or even where we want the majority of people to be. Don't hate the player, hate the game.

> The best part was the sheer technical talent in the US

This is just nonsense. What set the US apart is its mass and ecosystem.

What is just a good idea in any other country gets 10k users as a side project. Whereas in the US it can more easily become a multi-million VC backed 'homogenous' product for its massive population/market.

Popular tech is rarely the best at a technical level (even when the best is developed within the US), it just has more opportunity to acquire weight and momentum to become the default choice for a problem-space.

> This is just nonsense. What set the US apart is its mass and ecosystem.

That's only in post 2010s period, aka post apps and social media.

The boom in 80s/90s/2000s all of it kicked off from the deep technical know-how that rest of the planet just didn't have.

People can feel free to disagree, but they are really lying to themselves if they think anywhere else on the planet there was the sheer amount of tech talent as US until even 2010s.

Things have changed since and US is now largely valued because of it's large market and stuff, hence the off-shoring of so much of the technical and even advanced work.

But this was not the case before. With AI pushing the cost of building stuff to near 0 I don't see why you even need the US market to scale up an idea enough to prove viability. Sure the company might have to move to US, but there is no reason for it to be built in US anymore.

I went in trying to earn why Pi was such a good harness and did not learn anything. I’m really curious to learn why someone might consider it to be 10x better than Claude Code.

> Pi is Lindy ↗. It’s been around for a year and has arguably the biggest developer community across all harnesses. It’s here to stay.

BTW this isn’t what Lindy really means

Claude code is bloated af, super slow and doesn't give better results the most harnesses. Using codex, claude code and omp back to back makes it painfully clear.

The only place showing some numbers is https://frontierharness.org/ but it isn't really up to date, it's easy to test yourself though

I switched from Claude code to pi about two months ago and can’t imagine having to go back.

However, this is very much a “build what you need” type of harness. You have to be willing to use other people’s extensions or build/vibe your own before the harness is useful.

Why I find pi better:

- multi-provider/inference API support out of the box

- greater visibility into token flow/usage

- minimal system prompt

- no hand holding/decisions made for you on what your agents can/cannot do

- extremely customizable

- stable, minimal bugs in the base harness that I’ve experienced

- minimal feature bloat

It’s a great harness for someone who wants extreme control and adaptability. I don’t recommend it to my peers and colleagues who aren’t into tinkering.

> I don’t recommend it to my peers and colleagues who aren’t into tinkering.

For them/me, omp (https://omp.sh/) is excellent and batteries-included. Pi = Linux kernel, omp = Linux/GNU.

I have journeyed from Claude Code to OMP to Pi while working very hard to keep a super open mind, open to anything.

This from the post is why I landed on Pi

"What makes Pi the best fit here? Simple: it’s the only harness today that’s open source, vendor-agnostic, extensible, and minimal enough to not force opinions on you. It gives you full ownership and puts you in control. Being able to change the providers without switching apps alone is reason enough to adopt it, but the fact that it’s also so flexible makes it almost impossible for me to switch back."

I am building my own (intentionally simple) subagent orchestration framework inside Pi and am watching Pi Durable closely as I suspect it and Pi will unify before long and subagents are better supported in Durable. (Relatively uninformed opinion in this case, not using Durable yet.)

Edit-typo

Invoking the Lindy Effect to proclaim that something that’s a year old will last forever is pretty funny. What the Lindy effect actually predicts is that something that’s a year old is pretty likely to only last one more year before being replaced.

Tau is kind of taken as a name, BTW, by the Python harness from Huggingface. And I do agree Pi is great, but what makes it great is the back-end (pi-durable, pi-ai, etc.).

I've been working on https://github.com/rcarmo/piclaw for most of the year, as well as https://github.com/rcarmo/gi and https://github.com/rcarmo/go-ai (as ports of pi and pi-ai/durable/etc., plus 3 or 4 pi-durable add-ons for Obsidian and other desktop apps) and the internals of pi are a grand thing indeed, because it was designed from the outset to be able to extend itself (which many people miss, because many people confuse the harness with the TUI).

Why the downvotes? Are people completely oblivious to https://github.com/huggingface/tau ? That far removed from the fact that the harness is not the TUI?

I agree that pi is very good, don’t agree that it is the last harness (I have written 4 harnesses myself this year, fantastic experience.)

I mostly use local models. I use pi with the strongest models my 64g mac can handle, and for fast light work I use ‘pool’ with Laguna 2.1 XS 6 bit quant. I am happy :-)

VSCode is the last editor, pi is the last harness, blah, blah blah

Pi only relatively recently getting Pi Durable feels like the first time we could even start to entertain it as a long term option. Having memory, history feels like an incredibly low bar to cross.

I don't see much argument here. This feels like a variety of "worse is better".

> Separately, Pi is good enough that I don’t expect an alternative to be an order of magnitude better.

I fundamentally cannot take this seriously as an argument. For a number of ways, one of them being the focus on how many zeroes can be added. These are such encompassing personal spaces for us, connect person, agent, code and tools so intimately. The idea that oh, this is good enough, feels preposterous. As so many others have said, we are just getting started.

Pi itself is mostly blank slate, invites you to get started. To build something far better. I do think there's absolutely something to that idea, that the harness is sort of this start, a base, to build your agency atop. And I deeply think that your harness must be your to improve and iterate on, to adjust, that only those ecosystems that let people in will be of any long term interest and value, will be the site if innovation and building, the collectors of energy. The post shouts out DeepSeek Harness. I think OpenCode with its effect based core is also spitting some real fire too. As they ask on the Bluesky, "what else is like this?"

I don't love the attempt to clarify the term "harness" or the analogy.

"Engine" works well as the model. The car is the harness.

Pushing back further (etymologically): A horse is the model. The harness connects the horse to the coach/wagon, where the user sits. Multiple horses may be connected in parallel. Maybe oxen (different models) are connected for certain jobs, which require a harness appropriate for the workload.

I agree that harness, like agent, is a hype word and is often part of word soup. It's been at least a year now (an eternity in the AI wars) where the frontier model releases are just one factor we are using and weighing against competitors, and things I would put under the "harness" category matter as much or more.

>LLMs change everything around us, how we manipulate the world, and how we get the signals back.

This is sounds a little too grandiose for me. Our physical lives at least seem quite unchanged. If you don't produce code or papers for a living you might even succeed in not thinking about LLMs at all. The "signals" we're getting have certainly multiplied and increased in amplitude, though.

Regarding Pi: If hands-off code production is the future wouldn't a harness with subagents be the obvious choice? I personally enjoy minimalism in software, if only for aesthetic reasons, but even then: Can't you spin up your own bespoke version easily?

My guess is we see a cambrian explosion in harnesses. Pi may well be the substrate for lots of them. But as coding agents get better, everyone can have a version that fits their needs like a glove.

> everyone can have a version that fits their needs like a glove

That's kind of the point of Pi in the first place.

Is anthropic ok with using their models in non claude code harnesses?

Completely. You just use their api and you pay for it. The harness needs to speak the anthropic messages protocol or you use a proxy that does the translation for you.

I think working in a filesystem is a local maximum and making a harness that is agnostic of this might be a good idea.

(True enough: It's a local maximum that is hard to overcome. But we already see glimpses here and there such as anthropics scion: Why would an agent even check out the codebase if it could delegate testing, etc to a bunch of runners?)

i love pi and built my harness on top of it. but while checking error rates on tool calls, I noticed even opus 5.5 was making mistakes, like calling bash when the tool was named shell, or sending old_str or new_str instead of old_text or new_text (basically using the names of the tools in claude code despite tool definitions being different in my harness. i added aliases to fix it, but it changed how i think about harnesses now.

models are being heavily rl'd on the lab's harnesses, so eventually claude works better in claude code than any other harness.

i'm moving my weird customisations a layer above to acp or building on the claude sdk, especially since they now allow using your sub w it.

the labs are making it clear that they want control of the harness as well

I've seen many mentions of pi and omp, but I can't figure out what they bring to the table. Does one need to be in the 10x club? I don't have herds of agents, I have different "chats" for different project or search queries, on the desktop and CLI.

What will break pi is that everybody is going to think themselves clever by calling their extension/frontend/fork 'tau'.

> Sure, there are certain features [....] but I don’t see how those couldn’t become part of Pi eventually, either natively or via plugins.

No. I'm done with npm-style dependency sprawl and the supply chain nightmares. Done with plugins, extensions, marketplaces with "install at your own risk" disclaimers.

Give me a opinionated monolith by somebody who has something at stake and doesn't externalize these costs, i.e. a corporation.

Maybe once we live in a world with lean-verified wasm-sandboxed provable-secure software we can get back to installing software written by randos on the internet.

Nah, pi to other harnesses is like archos to other distros: if you like spending your time tweaking and customizing the tool itself rather than doing the work, it will be the perfect match, that, plus it’s turning into a cult-like, exactly like archos. OMP is too convoluted with features that also resemble an old linux distro (can’t remember its name, i think ultimate) where the creator tries to shove all features in there, i remember that distro had all programs preinstalled plus all themes and everything, same with omp, plus omp terminal ui is ugly and behave badly if you resized the window, it feels like a hacky project. What’s really needed in any good harness: multiple providers, efficient on tokens, load skills, easy export of sessions, fast with clean ui, and open to customization if you needed with addons.

I'm sure there are many more harnesses like Pi. Clai[0] (which I've written) is one of them. We also have hax[1] and aichat [2]. I made some additional comparisons here[3].

At the end of the day, it's quite easy to write together a bespoke agentic harness. So I'm sure there are countless additional ones. You don't really need a lib/open source one at all. Just make it write itself.

[0]: https://github.com/baalimago/clai [1]: https://github.com/OleksandrChekhovskyi/hax [2]: https://github.com/sigoden/aichat [3]: https://www.clai.chat/compare/

palgu.ai is the ultimate superharness :)

Yeah? Well, you know, that's just like uh, your opinion, man.

___

What is the point of this blog post, declaring a problem solved with.. uh.. authority taking from.. uh.. uuuuuh. Anyway.

Is pi good? I guess? The post does mention some of its qualities.

Is it a sensible take to extrapolate this from that? I do not think so.

Intellectually unstimulating.

this article is hyperbolic, but pi is very nice, if a little raw for people used to more professional commercial harnesses

i am vibing a harness extension on top of it for our students called catatui:

https://msu.github.io/catatui/

that works w/our free-for-students-and-staff internal models & provides some student-oriented features, and it has worked very well

Reply on news.ycombinator.com