I realised how use case dependent harness behaviour is when I tried to use my customised-for-a-side-project pi config at work and realised I needed to tweak it significantly to be useful - I would not be surprised if tools like Claude Code needing to be all things for all people is hurting their peak usefulness.
Having tried all the coding harnesses, I find that using Pi is exactly like using Emacs. For anything you want to build you can ask your agent and it will build it. There's tons of existing code to help you configure it. At the same time half the code is buggy, UI elements will try to overlap one another, and you'll periodically get crashes.
If you're willing to put in the work to master the learning curve and push through the issues, it can be a great tool: https://i.sstatic.net/7Cu9Z.jpg
Literally every AI project ends up getting monetized even (read: especially) if it starts as open source. It's only a matter of time before π's enshittification kicks in.
We built our first version of our in-app agent with opencode, but we're probably going to switch to Pi. Fundamentally, Pi is the shape of the thing that should exist, and it's awesome to see that it works. Very satisfying to my engineering brain. Been having fun with ohmypi as my personal harness, but going to switch over to pi soon and just mess around with it until it works the way I want. The idea of completely personalized software is so cool and now that it's finally possible I am having so much fun.
Pi has the best code out of all coding agents. The only problem is it’s not as fully featured and some of its modules should be factored out into their own code base.
I've had a lot of success running Pi on my server in headless mode and wrapping it in an XMPP client. This means I can talk to it wherever I can access XMPP (everywhere). It also mean agents can talk to each other when they need to. They've got a shared wiki they interact with and github issues as their todo list.
I am running several named pi instances in parallel in their own user account on NixOS, so they can install whatever they want in ephemeral shells and I never need to worry about their env. The agents can spin up new enabled XMPP agents if I request it, though for now I've only needed a few since I'm not doing too much in parallel.
My Pi is very vanilla, only my own XMPP wrapper and pi-subagents extension for anonymous subagents.
Using it primarily with Deepseek v4 Flash for chipping away at coding tasks or server maintainence while I'm AFK or in transit.
NixOS is the key to all of this, since agents can interact see the whole server config, make changes and run compile-time checks before actually deploying. It also means that even if they do mess up I can always revert.
I often find it reductive when people say "just tell Pi to build you an extension". Having used it as my one and only harness for a few months now, it's easy to get an extension, but hard to get a good one, that actually works well and helps.
My advice: focus on getting work done and slowly adapt Pi with small augmentations as you go. You can start getting work done on vanilla setup. When the right idea comes along, try it. Be ready to refine it, and most importantly, rollback the addition. I've rolled back a bunch.
Many "batteries" that are "included" come from speculative and half-baked ideas, from people who were excited about something at some point in their journey. In practice, those ideas may not bring the desired results, and their creator may've moved on already. So it's better to either learn very well established tools, or mold your own slowly.
For example, many automatic memory systems are not helpful. I built a small extension that asked me whether it should remember something (and write it down to a properly scoped SKILL or AGENTS file). Turned out I accepted less than 5% of suggestions. Most were useless one-offs that would pollute the context. Can't imagine how much crap would accumulate if I wasn't in the loop.
The biggest issue with Pi is that they don't have proper sandboxing with auto approval. Most solutions are third party and half baked. You have to either choose between just auto approval (no sandboxing), or a Claude/Codex style sandbox but manual approval.
I managed to get the sandbox working, and the option to break out of sandbox (with approval).
The network sandboxing doesn't work yet (I don't want malicious script to be execute and exfiltrate the data), I created an issue here: https://github.com/erichll/pi-packages/issues/1 - might submit a PR once I get a reply.
Honestly I think the thing I love most about Pi is that I'm not dealing with a tool with a thousand undiscovered features. I either pick and install what I want from the package ecosystem or I bottle up my own workflows into extensions as I find what works best for me.
The result is that the tool gradually morphs into the thing I need rather than me having to adapt myself to whatever new thing Anthropic or OpenAI comes up with.
I can also feel confident that the thing it becomes is what I actually need and not what maximizes token usage...
Aside from the minimal system prompt, how does it handle context better than other agents? It still has to send the system prompt (which includes AGENTS.md and skill definitions) along with the full conversation every request, no?
the compaction does not compact whole context, but keeps last ~20k tokens as is, I believe this helps a lot to model to not get confused what it is doing right now.
it also have soft/soft compaction limit, it tries to compact on turn boundary when possible. with combining with above this can get you about 35% more context (at least it looks like this with the sol)
codex when shell command is executed, will pull output with hard cap at max 30s, so for running compilation it will burn tokens without any benefit.
I have some tasks where agent will have to run some suite that can take over an hour, and codex burns about $20/h just waiting and reasoning every 30s "yep, that's still running". And what is going to happen after compaction, when whole context was just waiting? it will loose the plot and when I'm back it just does completely different thing that I asked it to do.
codex also have a bug, that opening refuses to resolve that adds your last steer after compaction, so imagine that you asked it to cleanup some tmp files or refactor/simplify something. it will do that again and again after each compaction, best case it just burns tokens and figures out, this is already done, or worse do it again and mess up everything and forget about it's task
i really hope OpenCode v2 with the Effect internals can ascertain and realize and lock on to being the best extensibility platform, on to having a core that serves not just itself, it's conceits of/as product, but a wider world. a street that wants to "find it's own uses for things".
it feels like there's still a strong layer of "for us" vs "for you" within opencode, that i hope, over time can get chewed away at. plugins to rebuild history, to re-title are just impossible, for example. none of these changes, these freedoms are hard to release. the patches i juggle are easy. but whether or not my agentic software serves as a substrate for my desire, or whether it allows me to augment agency: tis the question.
Dax (opencode lead) has such humble takes, is so forthwith about trying failing trying again on and on. about iterative improvement. and it feels like the guts are so in line to deliver, to allow such freedom now in OpenCode. but i don't see the product (anti-product) alignment, where opencode understands that it's competition isn't cc or codex, which can't and won't ever really compete, but pi, that the competition is to be the putty, to deliver the agency, to be a substrate. really hoping, because i love opencode, and these internals in v2 are sick.
the "devtools must be open sourced" debate comes screaming into the fore on this. it certainly argues similar to the post here: that it is minimalism, it is adaptability, programmability, it is directability that unlocks and unleashes us:
i don't even fully agree! today more than ever, why not cut a VS Code extension? why not cut some wild coop.nvim async extension that runs whatever subprocesses, talks to whatever system daemon? dream it up and do it; the llm's will cut through the mechanicals. but the core point, about finding software that doesn't obstruct, that accelerates the human agency: it's so Douglas Engelbart. to Augment Agency is so close akin to Augment Intellect, the grand passion for human interest engagement envolvement constructivism fucking-around-and-finding-out. and my how unhindered we can be now. if only our tools/systems/softwares let us be. here's to you, soft software!
Pi’s minimalism reveals a simple truth: the LLM is the core of any agent harness. Consequently, much of current harness tuning will become redundant(or even a hindrance) with next-gen models.
I found the nice part of Pi was the plugin system and the ability to hook into any part of the agentic loop. But minimalism is a net negative in my experience. What you really want is a harness that meets the expectations of the model and steers it in the right direction.
There are also a ton of small mechanical things a harness can handle that make the whole process much smoother. A really simple example is auto balancing parens. Even frontier models like Claude still struggle with this. Often the model will end up writing a python script to figure out where the mismatch is, and then generate a new version of the code. All of that simply wastes tokens and eats up context on a task that could've been accomplished completely mechanically.
The approach I took with dirge, is to put the model in a loop where it has clearly defined tasks, and the harness handles any repairs that can be done automatically. And I used Janet to provide a plugin system based on what Pi is doing. You get a batteries included experience out of the box, and you can customize it to fit a specific project using plugins if needed.
If someone has never used Pi, what's a good starting point where someone took Pi and started adding some extensions etc to get to that minimal-but-useful-and-productive-for-coding state? Is there such recipe/writeup?
Pi does one thing that I love, developing a tool that has minimalism where it's easily configurable with good documentation. The leads to new use cases that the the author(s) would have never dreamed of. The organic growth process of the Pi ecosystem has been fascinating to observe. It's one of the reasons why Pi has become one of my favorite coding agents to this day, flexible beyond personal uses and extensible to larger environments.
IMO, I view it more than a coding agent, it's a coding agent platform with powerful extensibility.
I like Pi, but I didn't end up using it. I tried OpenCode, Pi, Zed Editor and some NVIM packages. I wanted open source, featureful and easy to use. Specifically I wanted to easily edit the agent prompt.
I ended up on VS Code. I'm very critical of Microsoft generally, but VS Code is a very good editor and my favorite agent harness.
I think for indie hackers and people that build their own stack is great, but real scenario and people with money Enterprise likes the idea of batteries included.
I found this one a little bit better and they do support Extensions like Pi. But comes with all features like codex, claude code and it's open-source.
95 comments
[ 0.27 ms ] story [ 46.3 ms ] threadI realised how use case dependent harness behaviour is when I tried to use my customised-for-a-side-project pi config at work and realised I needed to tweak it significantly to be useful - I would not be surprised if tools like Claude Code needing to be all things for all people is hurting their peak usefulness.
If you're willing to put in the work to master the learning curve and push through the issues, it can be a great tool: https://i.sstatic.net/7Cu9Z.jpg
I am running several named pi instances in parallel in their own user account on NixOS, so they can install whatever they want in ephemeral shells and I never need to worry about their env. The agents can spin up new enabled XMPP agents if I request it, though for now I've only needed a few since I'm not doing too much in parallel.
My Pi is very vanilla, only my own XMPP wrapper and pi-subagents extension for anonymous subagents.
Using it primarily with Deepseek v4 Flash for chipping away at coding tasks or server maintainence while I'm AFK or in transit.
NixOS is the key to all of this, since agents can interact see the whole server config, make changes and run compile-time checks before actually deploying. It also means that even if they do mess up I can always revert.
My advice: focus on getting work done and slowly adapt Pi with small augmentations as you go. You can start getting work done on vanilla setup. When the right idea comes along, try it. Be ready to refine it, and most importantly, rollback the addition. I've rolled back a bunch.
Many "batteries" that are "included" come from speculative and half-baked ideas, from people who were excited about something at some point in their journey. In practice, those ideas may not bring the desired results, and their creator may've moved on already. So it's better to either learn very well established tools, or mold your own slowly.
For example, many automatic memory systems are not helpful. I built a small extension that asked me whether it should remember something (and write it down to a properly scoped SKILL or AGENTS file). Turned out I accepted less than 5% of suggestions. Most were useless one-offs that would pollute the context. Can't imagine how much crap would accumulate if I wasn't in the loop.
It is actually a combination of 3 extensions
I wrote a piece about this here: https://ptgamr.substack.com/p/a-pi-setup-with-permission-san...
It is a combination of 3 extensions.
I managed to get the sandbox working, and the option to break out of sandbox (with approval).
The network sandboxing doesn't work yet (I don't want malicious script to be execute and exfiltrate the data), I created an issue here: https://github.com/erichll/pi-packages/issues/1 - might submit a PR once I get a reply.
The result is that the tool gradually morphs into the thing I need rather than me having to adapt myself to whatever new thing Anthropic or OpenAI comes up with.
I can also feel confident that the thing it becomes is what I actually need and not what maximizes token usage...
it also have soft/soft compaction limit, it tries to compact on turn boundary when possible. with combining with above this can get you about 35% more context (at least it looks like this with the sol)
codex when shell command is executed, will pull output with hard cap at max 30s, so for running compilation it will burn tokens without any benefit.
I have some tasks where agent will have to run some suite that can take over an hour, and codex burns about $20/h just waiting and reasoning every 30s "yep, that's still running". And what is going to happen after compaction, when whole context was just waiting? it will loose the plot and when I'm back it just does completely different thing that I asked it to do.
codex also have a bug, that opening refuses to resolve that adds your last steer after compaction, so imagine that you asked it to cleanup some tmp files or refactor/simplify something. it will do that again and again after each compaction, best case it just burns tokens and figures out, this is already done, or worse do it again and mess up everything and forget about it's task
it feels like there's still a strong layer of "for us" vs "for you" within opencode, that i hope, over time can get chewed away at. plugins to rebuild history, to re-title are just impossible, for example. none of these changes, these freedoms are hard to release. the patches i juggle are easy. but whether or not my agentic software serves as a substrate for my desire, or whether it allows me to augment agency: tis the question.
Dax (opencode lead) has such humble takes, is so forthwith about trying failing trying again on and on. about iterative improvement. and it feels like the guts are so in line to deliver, to allow such freedom now in OpenCode. but i don't see the product (anti-product) alignment, where opencode understands that it's competition isn't cc or codex, which can't and won't ever really compete, but pi, that the competition is to be the putty, to deliver the agency, to be a substrate. really hoping, because i love opencode, and these internals in v2 are sick.
the "devtools must be open sourced" debate comes screaming into the fore on this. it certainly argues similar to the post here: that it is minimalism, it is adaptability, programmability, it is directability that unlocks and unleashes us:
> Imagine the convoluted misery it would be trying to plug that into the VS Code extensions API! Or trying to get it into vimdiff. It would certainly be possible, but the machinery to start pre-processing the commits as soon as they appear would be nigh-on impossible. - https://blog.exe.dev/devtools-must-be-open-source https://news.ycombinator.com/item?id=49156111
i don't even fully agree! today more than ever, why not cut a VS Code extension? why not cut some wild coop.nvim async extension that runs whatever subprocesses, talks to whatever system daemon? dream it up and do it; the llm's will cut through the mechanicals. but the core point, about finding software that doesn't obstruct, that accelerates the human agency: it's so Douglas Engelbart. to Augment Agency is so close akin to Augment Intellect, the grand passion for human interest engagement envolvement constructivism fucking-around-and-finding-out. and my how unhindered we can be now. if only our tools/systems/softwares let us be. here's to you, soft software!
There are also a ton of small mechanical things a harness can handle that make the whole process much smoother. A really simple example is auto balancing parens. Even frontier models like Claude still struggle with this. Often the model will end up writing a python script to figure out where the mismatch is, and then generate a new version of the code. All of that simply wastes tokens and eats up context on a task that could've been accomplished completely mechanically.
The approach I took with dirge, is to put the model in a loop where it has clearly defined tasks, and the harness handles any repairs that can be done automatically. And I used Janet to provide a plugin system based on what Pi is doing. You get a batteries included experience out of the box, and you can customize it to fit a specific project using plugins if needed.
https://yogthos.net/posts/2026-06-08-dirge-code.html
at least with GPT 5.6 Sol fwiw
https://smolenv.com/t/nested-template-includes-60636/
sh is all you need
IMO, I view it more than a coding agent, it's a coding agent platform with powerful extensibility.
I ended up on VS Code. I'm very critical of Microsoft generally, but VS Code is a very good editor and my favorite agent harness.
For headless, Pi might be the way.
#teampi
I found this one a little bit better and they do support Extensions like Pi. But comes with all features like codex, claude code and it's open-source.
https://github.com/autohandai/code-cli