Descriptions:
AI Jason walks through a practical workflow for cutting AI token consumption by using tmux — the terminal multiplexer — to coordinate persistent sessions across multiple coding agents. The approach builds on a pattern that Claude Code and Devin have both highlighted: using a powerful model like Claude Fable 5 as an orchestrator/planner while delegating execution to smaller, cheaper models like Sonnet 5, which costs roughly 20% of Fable 5 but performs comparably to older top-tier models.
The video explains why traditional sub-agent spawning wastes tokens — each new session starts cold with no context — while “sidekick” or persistent agent sessions reuse cached context at approximately 10% of the standard input token cost. Jason shows how to configure a coordinator-executor pattern in a claude.md delegation rules file, then extends it beyond Claude Code to integrate Codex CLI, Gemini CLI, and other third-party coding agents through tmux’s ability to write to and read from arbitrary terminal windows programmatically.
Viewers get a live demonstration of spawning parallel agent sessions, sending follow-up messages to existing sessions, and capturing terminal output to pass results back to the orchestrating agent. Devin’s recently announced Fusion harness — which Devin claims achieves similar performance to Fable 5 at 35% lower cost — is used as a benchmark reference throughout. The approach is tool-agnostic and applicable to any developer running multi-agent coding workflows who wants to reduce spend without sacrificing output quality.
📺 Source: AI Jason · Published July 20, 2026
🏷️ Format: Tutorial Demo
