Tmux + Fable = Cut 35% less token

2026-07-20 · Source: AI Jason · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Intermediate, long

Summary

This analysis explores strategies for optimizing Large Language Model (LLM) usage and cost efficiency, particularly with models like Fable 5 and Sonnet 5. It advocates for an orchestration approach where a powerful model, such as Fable 5, acts as an advisor or planner, while smaller, more cost-effective models like Sonnet 5 serve as executors. Devin Fusion, for instance, achieves similar performance to Fable 5 with 35% lower cost by combining Fable 5 orchestration with Sonnet 5 workers. A key technique is the "sidekick model," which utilizes persistent agent sessions to maintain context and reduce token expenditure by leveraging cached tokens. The article demonstrates implementing delegation rules in Cloud Code for this setup and extends multi-agent orchestration to other platforms via a Codex plugin or a universal Tmux-based method. It also introduces an "open agent teams" skill for managing diverse agents and highlights premium orchestration tools like Herd and Orca, noting Orca's intuitive UI and features for tracking agent sessions and token usage.

Key takeaway

For AI Engineers managing LLM costs in multi-agent systems, prioritize an orchestration strategy where powerful models like Fable 5 plan, and smaller, cheaper models like Sonnet 5 execute. Implement persistent agent sessions to leverage cached tokens, drastically cutting expenses. You should configure delegation rules in your `cloud.md` or explore tools like Orca for intuitive, universal agent orchestration, ensuring efficient resource utilization and faster task completion.

Key insights

Orchestrating powerful LLMs as planners with smaller models as executors significantly reduces cost and improves efficiency.

Principles

• Delegate complex planning to powerful LLMs, execution to smaller ones.
• Utilize persistent agent sessions to cache context and reduce token cost.
• Universal agent orchestration can be achieved via terminal multiplexers like Tmux.

Method

Implement delegation rules in `cloud.md` for coordinator/executor roles. Use persistent agent sessions (sidekicks) to maintain context. For universal orchestration, use Tmux to control and communicate with diverse agent CLI sessions.

In practice

• Configure `cloud.md` with delegation rules for Fable 5 (coordinator) and Sonnet 5 (executor).
• Install the Codex plugin CC tool to delegate tasks to Codex agents from Cloud Code.
• Use the `open agent teams` skill to orchestrate multiple agents via Tmux for any terminal.

Topics

• LLM Cost Optimization
• Multi-Agent Systems
• Agent Orchestration
• Persistent Sessions
• Cloud Code
• Tmux
• Orca

Best for: AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Agentic AI: How to Save on Tokens — Optimizing LLM agent costs requires strategic token reuse, context minimization, intelligent model routing, and diligent context cleaning.
• Anthropic's fix for Fable 5's high cost is turning it into a manager that delegates to Sonnet 5 — High-cost LLMs can serve as efficient planners or advisors for cheaper execution models.
• The Harness Effect: How Orchestration Beats Token Maxing on Your AI Bill — The orchestration layer, not the specific LLM, primarily dictates the cost of agentic AI systems.
• Mythos is here, it’s time to start tokenmaxxing — Maximizing subsidized AI model usage requires strategic rate limit management and multi-agent orchestration for complex tasks.
• Remember, Don't Re-read: Stateful ReAct Agents for Token-Efficient Autonomous Experimentation — Stateful ReAct agents significantly reduce token costs in autonomous LLM experimentation by maintaining context across iterations.
• Companies Are Making Claude and Codex Talk Like Cavemen to Stop AI’s Soaring Costs — Companies are simplifying AI outputs to reduce token costs, a direct response to soaring LLM expenditure.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Jason.