Technical Context
I love news like this not for the hype, but for its practicality. It finally discusses a real setup where Claude, Codex, and Gemini run as high-level CLI orchestrators, with Telegram acting as a simple entry point for tasks.
Essentially, the scheme is simple: I send a task to a bot, the orchestrator breaks it into subtasks, spins up the necessary CLI sessions, and collects the result. For AI automation, this is no longer a toy but a viable working architecture, as long as you don't try to run everything without limits.
I've looked into the available patterns, and the picture is quite clear. Claude is often used as a coordinator, Codex as a strong code executor, and Gemini as a dedicated worker for a narrow role. VS Code and Copilot have started to officially highlight such scenarios, while the community has long been gluing this together with shell scripts, markdown instructions, and file-based state.
The most practical point here isn't the choice of brand, but the execution model. A cold start provides a clean context and fewer surprises but consumes time for each launch. A hot start is faster, especially if tasks arrive in a series from Telegram, but then context drift begins, and the agent starts remembering irrelevant information.
I had to smile at the 10-agent limit on a Mac. Yes, it sounds like a normal engineering experiment until your laptop starts buzzing like a drone. In practice, around this threshold, you start seeing CPU spikes, memory pressure, IDE slowdowns, and an annoying delay in launching new CLI processes.
What This Means for Business and Automation
Realistically, the winners are teams that need a pipeline rather than a single “smart chat”: decomposition, parallel implementation, review, and response assembly. Here, artificial intelligence integration into Telegram or an internal chat really saves time because employees write tasks in a familiar interface instead of juggling five different tools.
The losers are those who try to keep this entire menagerie running locally on one Mac and call it production. I would keep one hot orchestrator locally and move the heavy workers to a Linux box or the cloud. Otherwise, the cost of chaos quickly negates any gains.
At Nahornyi AI Lab, we solve these kinds of problems at the architectural level: where to keep hot sessions, what to launch cold, how to limit parallelism, and where to offload executors. If your processes are already bottlenecked by manual coordination, my team and I can build an AI solution development tailored to your actual task flow, without the circus of ten overheated agents on a laptop.