Brief summary of previous situation..
Purchased M4 32GB to run openclaw on Mac Mini. I tried installing a local LLM, but it lagged even on the 30B model, so I gave up immediately.
And on the weekend, I switched to running it using commercial AI APIs.
After preparing OpenAI, Claude, and Gemini API keys, I attempted to implement an orchestrator by configuring the agents that exist in openclaw for each agent. This is a simple account of that.
An Orchestrator automates the work of AI agents, coordinates them in proper order, and efficiently manages the entire workflow. It's a term that's become quite popular recently because it integrates and manages interactions between systems like a conductor coordinating instruments, thereby increasing operational efficiency.
What you need to understand here is that I'm configuring this Mac Mini for miscellaneous tasks, so I'm not using the latest models.
I want to organize, analyze, and understand miscellaneous tasks, receive essential and organized information, and use it as a reference or direction when working. In other words, I want to use it as a support assistant.
(I want to receive first-processed information or organized content and minimize unnecessary information.)
And an orchestrator isn't necessarily always good. Simple tasks might be faster and better handled by a single agent. I intend to distribute tasks based on my belief that an orchestrator is more efficient when there's a large workload.
The final structure I'm pursuing is this:
The generated agents are PM, Orchestrator, Planner, Backend, Frontend, Designer, Crawler, Summary, Collector, Data, and Reviewer. I initially assigned gpt-5-mini, gemini-2.5-flash, and claude-sonnet-4-5 models to each according to their respective roles. And when I tested, it seemed to output proper results.
This is a lie. When I checked the active agents, I found that the PM alone was processing everything while appearing to have divided the work.
It wasn't actually calling or executing other agents.
I wanted the PM to execute agents as children, then send the results back to the PM, organize them, and report back to the user, but this part wasn't working.
Currently, openclaw has a structure that runs each agent independently, and it doesn't support this kind of orchestrator.
So I spent two days solving this problem, and it was quite difficult.
(Of course, I didn't do most of it—codex did most of the work. Haha, codex is great.)
The key parts I handled were:
PM/Agent Response Handling
Session Control (like sessions_spawn, sessions_yield)
Passing Subagent Completion Events to Parent
Assembling Messages to Display in Gateway and TUI
These are the parts I resolved by modifying openclaw's core file (reply-Bm8VrLQh.js).
Handling the sessions_yield part was the most difficult. (codex solved it after running alone for 1.5 hours)
After modifications, I successfully ran each agent in multi-threaded mode, aggregated the results, and returned the report to the PM.
However, I discovered that it was using the default AI model instead of the specified models..
After re-modifications, each agent's designated model operates and produces results to report to the PM..
Naturally, delegating tasks to each agent and receiving results back takes more time.
It takes roughly 30 seconds to 1 minute.. I think I need to verify and tune whether the results are better.
This is the usage I confirmed from the dashboard. Even with low-cost models, the cost seems to be higher than expected..
Though I tested without restraint, it seems to have cost around $0.9 over two days.
(openclaw passes along additional information besides the query, so it tends to consume many more tokens.)
The memory usage during operation reached 10-16GB.
For now, I've briefly written down what I struggled with over the weekend.
If there are those who want more information, I'll write it in detail again.
