Token Efficiency (www.june.kim)

A recent conceptual piece, "Token Efficiency," frames token budgeting like RAM and stack memory management for LLM-based agents, proposing practical heuristics for minimizing context pollution and improving result quality. It contrasts persistent MCP (multi-context processing) servers—which occupy a large share of an agent’s initial context window—with on-demand "skills" that only consume context when called. The author visualizes how tokens accumulate across context windows during conversations, describes "compacting" (manual cleanup) to mitigate context rot, and argues that spawning ephemeral subagents for focused tasks (e.g., research) prevents pollution of the main agent’s context even if it doesn’t speed execution. For practitioners building multi-agent or skill-based systems, the takeaway is actionable: prefer specialized, ephemeral subagents invoked as function-like calls when you only need final answers, and reserve MCP/stateful components for persistent needs. This is presented as a meta-cognitive optimization—planning errors that let irrelevant context accumulate can cost more in debugging than the original planning step—so designing for token efficiency improves model output quality, lowers inference/context costs, and simplifies reasoning about agent behavior. The framework impacts architecting LLM agents, cost optimization, and multi-agent decomposition strategies in production systems.