For the last several days, I have made my AI Agents call each other. Claude Code, Codex, and Gemini. After several agentic improvements, the root prompt does the following steps:
- Instructs an agent to copy and adjust the prompt for their task
- Instructs to use the standard
run-agent.shstart agents, log outputs - Clearly states that the AI Agent should use the script, not the embedded feature
- Starts the agent at any working folder
- One more secret ingredient to make it work, or two
One AI Agent controls the flow of a fleet of other agents. It is much better than just a bash loop, and I like how it performs and adapts to the task given.
Today:
- Deep research on a product topic was conducted by 16+ agents
- Code reading of a big monorepo, so the agent could understand how to implement the REST API client
Claude Code calling multiple agents on the JetBrains Space reposiotry to mine knowledge from my repositories
There is more
- Agent processes never ask me, and bother much less
- I manage to make a root AI Agent run for hours unattended
- It keeps the context small for each and avoid context rot
- One can manage 3-5 such sessions, and I want to grow this number
- The baseline is still the https://jonnyzzz.com/RLM.md, which I based on the outcomes of my multiple experiments
- Look what I’ve done MCP Steroid
What’s next? One part say – do measurements. I will keep experimenting at the first place, and I need a partnership to set it up in a more scientific manner.
This work builds directly on the success of my previous experiment where 16 AI Agents Fixed Our Documentation Problem and relies heavily on the Recursive Language Model (RLM) methodology I established earlier this year.
The fleet is busy workin’
Happy Friday!