Hi HN! I believe agents running on markdowns was the first step but instead of making the filesystem forkable/clonable I believe event sourcing on a real database is a better concept for agents to understand all the work that has been done, learn from its mistakes and self-improve. Happy to answer questions!
I love the concept to use different harness frameworks in qm but a true multiplayer harness needs to support other agents and any MCP clients, including Cowork.
Making agents multiplayer is mostly a context problem. You could be using ChatGPT or a Slack bot, or a web interface and the agent needs to know you, your conversations in Cowork etc. so it can enable multi channel collaboration with your agents and your colleagues. We're working on it at https://lobu.ai
why pi over opencode? earnestly curious, trying to figure out what open solution people are consolidating on. (codex is also pseudo-open but contributions closed and nice)
pi is the neovim of agentic harnesses, its barebones and extremely configurable. if you're the sort of person who likes that sort of things its a forever product, nothing is going to displace it because you have full control.
opencode builds a lot more in, which is better if you dont want to fiddle with config.
nice. i had thought the consensus had moved pretty firmly towards pi, so i was surprised to see Thinking Machines demoing their new model Inkling in OpenCode. wondering if they are previewing an acquisition
Agree, after spending too much time and tokens configuring Pi and adding extensions to match other harnesses, I switched to OpenCode and left the Pi customization circle jerk. I have other things to do and IMHO harness engineers should do the harness engineering, I don’t want to waste tokens and time to build and benchmark extensions. Pi is great, but would be better with a set of official, trustworthy and efficient extensions, and opt-in to enable it.
Most of my harness experience is with Claude Code and Pi, a little bit of OpenCode.
I like how quick and snappy Pi is, it feels like a minimal harness, just enough to manage the agent and get out of the way. Earlier models also seemed to have an easier time working with the tools, e.g. GPT-OSS-20B is about a year old and had no trouble in Pi.
I tried OpenCode but didn't particular like it as a Claude Code user, that is the main reason I switched to Pi. The reason I am sticking is how simple it is to extend it. I moved from Claude Code to Pi and within 2 hours (and the help of Claude Code) I have a setup that matches Claude Code and is even better for my setup.
Things I've added:
1. Built my own AI judge for 'auto' mode that matches my setup.
2. /plan /go for planning and executing.
3. /flow for A-Z setups. That includes planning, executing, testing and shipping.
4. /deep-research a multi fan-out setup for researching a topic.
5. My own sub agents.
6. A TaskCreate/Update/List setup.
7. Monitors.
8. BashOutput / KillShell.
9. Proper notifications with Notify that uses macOS banner and work.
10. Spawn tool that triggers multiple sub agents.
11. A bridge between signal to use Pi remotely.
Yes a lot of these things is something that was already in Claude Code but now I don't have to use Claude Code and I can customize it to fit me exactly.
I imagine because they want to support plugins, and plugins in compiled language are a lot less natural than plugins in languages like TypeScript or Python.
They play better with statically typed languages, not compiled ones in particular. Rust's typing is stricter than Typescript though so that probably helps.
Not really because you're not building a database or GUI app where using native elements & data structures help a lot with memory pressure.
TUI renderer is the one using the memory heavily so your terminal takes the heavy lifing. If you're managing the buffers and out-of-screen context good enough, Typescript can be pretty efficient.
I love opencode but it chews through memory on my 64gb MacBook Pro. Can’t have too many long running sessions because the memory use just slowly creeps up.
It’s not about the terminal at all which as you noted accounts for minimal isage. It’s all the internal chat and history and everything else the agent tracks - all of which are smallish (and largish) strings allocated on the heap.
I don’t have the same issues with rust based tuis.
I'm not sure why I'm being downvoted but I didn't mean it in a negative way.
For such announcement, I would expect them to give me clues on when I should use this model and in which cases it's the best one.
The benchmarks that they share doesn't indicate that it's cheaper to run than other models, or can fit in my local machine, or excels in a specific vertical.
After reading the comments here and X, I can see it being the top-3 multi-modal open-source model though.
That's why we built https://lobu.ai which is a multi-tenant implementation for proactive agents like OpenClaw. It has entity based memory for building the org context layer and every channel/user gets its isolated container.
The article lets itself down when it reveals it's had some marketing "polish" added: "where certain resources (like an AI model instance) are stored"
They could have picked something far more concrete (for Cloudflare's business and the type of services they deploy) than "AI model" there, which is clearly picked just to make it seem 'current' and on-trend.
That, or the model writing it already had "You are an AI model instance" sitting at the beginning of its context ready to pull from. Telling it it's an AI pushes it to make text about AI.
Filesystem is best when there is a single writer and many readers.
If you have bunch of files and don't have any structure, yes filesystem is great but the moment when you need consistency & performance (which you need sooner rather than later) use databases.
Investing early doesn't hurt when you build a product that you know will have many writers.
It's interesting to hear that 60% of PRs at Anthropic is created by the Slack bot. While building a Slack bot is easy, making it look like an AI teammate is pretty hard. This is exactly the gap we’re working on with Lobu.ai — disclosure: I’m the founder.
The hard problem is giving a shared agent durable organizational memory and a real isolated environment where it can safely access company systems and perform work. The agents need a durable log of what everybody at the company is doing, prevent data leaks with proper access control and isolate the runtime to give everybody both private & shared space.
It’s also not tied to Claude or Slack. We see Slack as one interface and the models as part of the harness. It's usually better to combine multiple providers to review the work.
It's not built on OpenClaw, it uses Pi the harness that powers OpenClaw. We were using the GTM for multi-tenant OpenClaw when there was hype but quickly moved away from it as OpenClaw is a nightmare in terms of security.
reply