I have a specific use case (financial analysis), that is at the edge of what is possible with this models (accuracy wise).
Gemini 2 was the beginning, you could see this technology could be helpful in this specific analysis but plenty of errors (not unlike a junior analyst). Gemini 2.5 flash was great actually useable, errors made were consistent.
This is where it gets interesting, I could add additional points to my system prompt, yes it would fix those errors but it would degrade the answer elsewhere, often it wouldn't be incorrect but merely much simpler less nuanced and less clever.
This is where multi-agents helped it actually meant the prompt can be broken down so that answers remain "clever". There is a big con to this, it is slow, slow to the point that I chose to stick with a single prompt (the request didn't work well operating in parallel as the other prompt surfaced factors for it to consider).
However Gemini 3 flash is now smart enough that I'd now consider my financial analysis solved. All with one prompt.
The quality of the output depends more on the underlying LLM. GLM 4.7 isn't going to beat Opus but Opus with an orchestra seems to be faster and perhaps marginally better than with a more linear approach.
Ofcourse this burns a lot of tokens but with a cheap subscription like z. ai or with a corporate budget does it really matter?
- there was a CLI based on OpenCode, that depended on extracting the auth token used by Claude Code
- Anthropic found a way to prevent this
- This person (or AI) reimplemented similar workflows using custom commands inside Claude Code instead
It appears that Anthropic is totally in their right to do this. Third party software is meant to use the API instead. It sucks that it's so much more expensive, but it's their decision to make,
Meanwhile, this seems like a good approach, both sides win. The project page would benefit from skipping all the drama and focusing on being a helpful intro.
I see "Multi Agent Orchestration", but, scrolling through this I still have no idea what I'm looking at.
I actually tried this few days back before the Claude Code EULA reinforcement, I went through the same thing.
1. I honestly had a hard time parsing what this is supposed to do or provide over standard opencode setup from the readme. It is rather long-winded and have a lot of bombastic claims but doesnt really explain what it does.
2. Regardless, the claims are pretty enticing. Because I was in experiment mode, and I already had a VM running to try out some other stuff, I gave it a try
3. From what I can tell, its basically a set of configs and plugins to make opencode behave a certain way. Kinda like how lazyvim/astronvim are to neovim.
4. But for all its claims, it had a lot of issues - the setups are rather brittle and was hard to get working out of the box (this is from someone who is pretty comfortable tinkering with vim configs), when I managed to get it working (at least I think its working), its kinda meh? It uses up way more tokens than the default opencode, for worse (or at less consistent) results.
5, FWIW, I dont find the multi/sub-agent workflow to be all that useful for most tasks, or at the very least its still very early IMO, kinda like the function calling phase of chatgpt to really be useful.
6. I was actually able to grok most of Steve Yegge's gastown post from the other day. He made-up a lot of terms that I think made things even more confusing, but I was able to recognize many of the concepts as things that I also had thought of them in a "it would be cool if we can do X/Y/Z" manner. Not with this project.
TBH, at this point im not sure if I'm using it wrong or am I missing something, or this is just how people market their projects in the age of LLM.
edit: what I tried the other day was the code-yeongyu/oh-my-opencode, not this (fork?) project
If you're hiring a consultancy or a pile of freelancers it's a bit different, but the question here would make me believe you don't trust their capability to start and I would be looking for teams that better align with what you expect as their outputs.
Specifically then, has anyone experienced a step change from these types of multi-agent tools, above the base experience of using Claude Code. Or is this something that just looks the part?
Reading through the code it gives the impression it's teasing out more power from the LLMs. The structure looks like an advanced use of Claude Code with correct feeling abstractions.
I dont understand how you can say that its a bad thing to ask an IT professional community like HN for advice?
I assume you were born with all the knowledge of the world?
