28 TOOL DEFINITIONS! — Cutting AI Agent Costs with Sub-Agent Splitting

This article is an AI-assisted translation of a Japanese technical article. Introduction I'm building a personal AI agent called TONaRi ("tonari" means "next to" in Japanese — named with the idea of an AI that stands next to you and supports your daily life). It's built with Strands Agents + Amazon Bedrock AgentCore, with a VRM-powered 3D avatar frontend using AITuberKit. As I kept adding tools to make my personal AI agent more useful for daily tasks, the input tokens per API call ballooned — and so did the cost. It's lower now, but the projection was heading toward $120/month In this article, I'll walk through the input token bloat problem caused by too many tools and how I tackled it by splitting into sub-agents. Architecture Overview Here's a high-level look at TONaRi's architecture: Frontend (Next.js + VRM 3D Avatar) → Next.js API Route → AgentCore Runtime (Strands Agent) → AgentCore Gateway → Lambda functions (tools) → AgentCore Memory (STM/LTM) The agent runs as a container deplo