Show HN: A lightweight way to make agents talk without paying for API usage

It has been reported that developer Juan Pablo A. published a simple — almost cheeky — workflow for making LLM-based agents talk to each other without integrating APIs or adding SDKs. The trick: use the CLIs that come with the models, resume previous sessions, and reuse the subscription plans you already pay for. The post allegedly promises you can avoid additional per-call API charges while running quick multi-agent experiments. Tempting, right?

The pattern

The core idea is surprisingly low-tech. One agent writes or reviews a draft, then invokes another agent from the command line using a “resume” mode so the conversation continues instead of starting from scratch. Examples in the post include commands like codex exec resume --last "prompt" and gemini -r latest -p "prompt". The author keeps conventions in a shared Claude memory file so agents know how to call each other consistently — think of it as a lightweight protocol rather than a full framework.

Why you might care (and what to watch out for)

This lets you stitch together different vendors for varied perspectives: ask Claude to draft, Codex to review, Gemini to offer a third take. It’s great for rapid experimentation and avoids extra dependencies. But there’s a catch: visibility and observability are limited. You can make agents talk, but it can be hard to inspect the full history, trace decisions, or manage permissions when one agent autonomously invokes another. Not exactly a production-ready pipeline — more like a scavenger’s delight for tinkerers.

When to reach for tmux

If you want more transparency, the author suggests using tmux to run each agent in its own pane or session. That approach requires more setup but gives you a live view of what each agent is doing and makes it easier to capture outputs. Commands and examples in the post show how to create isolated tmux sockets and send prompts programmatically. Handy. Practical. A little hacky. But for anyone tired of copy‑paste handoffs between models, it’s a neat trick to keep in your toolbox.

Sources: juanpabloaj.com, Hacker News