I Built a Coding Agent With the Manual Tool-Use Loop. Here's What It Taught Me

Short summary

Manual agent loops give developers control over the agentic pattern that SDKs typically hide. This post walks through a working TypeScript implementation showing how to add approval gates before destructive actions, plus practical guidance on turn limits, logging, and conditional execution. Understanding the loop fundamentally changes how you debug agents—from 'why did magic break?' to 'I can trace the exact decision tree.'

• •Agents are fundamentally while loops: request → response → tool_use → result → repeat
• •Manual loops enable approval gates, logging, turn limits, and conditional execution between model decision and action
• •Trade-off: manual loop requires more code but gives full control; use SDKs when no intervention gates needed

Generated with AI, which can make mistakes.