Current state of AI orchestration

As a follow up to my post about OpenCode, I am also interested in AI orchestration tools, like PaperClip or Gastown.

In this post I am talking about the integration of OpenCode to Paperclip as orchestration tool. It can be integrated into PaperClip - a “software factory for AI agents”.

PaperClip

The general idea sounds as crazy as it gets. A single agent is created as a CEO and has the ability to “hire” (aka create) other agents to work with him. This way, an Agent can specialize on a specific task. In theory.

The fun thing is, that this does not work at all in the current setup. The agents are communicating with each other through tickets, something is failing, nothing productive happens.

It can also very well be that I did not invest enough, as I did only run GPT-OSS 120b as a model with it. Interestingly, the mistralai-mistral-small-4-119b-2603 model does not work well for programming and tool calling - even though it was meant to be good at this.

An overview of the user interface can be seen below. Agents can hire new personas. Each being another instance of openai/codex/claude/opencode. Thats a great tool to burn through tokens.

AI slop - What a time to be alive

Of course, as an “agentic” tool, paperclip is also mainly developed by LLM tools.

When configuring Paperclip, the docs are often not helpful as they change all the time, so that links to them often not working. Links within the docs are also not consistent, which makes it hard to find what you need.

When configuring OpenCode as an agent to be used within Paperclip, the paperclip config environment is not used after creating an agent: https://github.com/paperclipai/paperclip/issues/5507

Workspace detection does not work with paperclip: https://github.com/paperclipai/paperclip/issues/469

And various other issues I have run into, which are also not fixed after a few weeks. We will see where this is going. As Open-Weights models are getting stronger, they will be cheaper to run and will not be stuck in loops as they are currently. For hosted frontier models, these tasks are far to expensive, as the whole “company” does not really move forward together.

Wrap up

Related post with similar experience for Gas Town: https://www.dolthub.com/blog/2026-01-15-a-day-in-gas-town/#did-it-work

With https://github.com/dtnewman/burn-baby-burn you generally burn your tokens more efficiently and do not receive anything from it either.

Lets take a look back to it in 6 months and see if we solved layer 2 AI.