I have been experimenting with several coding-agent harnesses, but the useful improvement did not come from running more agents. It came from giving one agent responsibility for coordination.
I usually brainstorm with a Codex manager first. When a task becomes concrete, it starts an executor in the harness that fits the work and hands over the objective, constraints, relevant files, and expected validation. I monitor progress through the manager, although I still enter the executor session directly when it needs steering.
The two pieces that made cross-harness delegation reliable were reusable skills and shared context. Skills encode the development process. Files hold current project artifacts, while a separate memory layer holds durable decisions, conventions, and verified lessons.
I also keep the same setup running on a VPS with Telegram connected, so I can delegate an idea from my phone and return to it later.
I wrote up the full workflow here: https://codeaholicguy.com/2026/08/19/how-i-run-ai-coding-agents-as-a-team-with-ai-devkit/
I am curious how other people handle context transfer and review when several coding agents are active. Do you use a manager, or supervise each session yourself?