You Should Write An Agent

Source: fly.io Date: 2025-12-10 URL: https://fly.io/blog/everyone-write-an-agent/

Summary

Essay by Thomas Ptacek arguing that building a basic LLM agent is the only way to develop informed opinions about the technology — “you only think you understand how a bicycle works, until you learn to ride one.” The post shows a functional agent in ~15 lines of Python, extends it with tool-calling via JSON schema, and identifies “context engineering” (managing token limits across agent calls) as a legitimate unsolved programming problem worth serious attention.

Implications

Builder-cultivation pattern. The essay is a useful onramp signal: Fly.io is explicitly cultivating a builder audience that will run agents on their infrastructure. The framing that agents are “deceptively simple” at the core but that context engineering is hard and real points toward where tooling will develop. For the radar, this is part of a cluster of fly.io signals (Sprites, MCP integration, “our best customers are now robots”) that together describe a deliberate platform play for the agentic workload market. The “specialized sub-agents and multi-loop architectures” line points toward the multi-agent coordination problem space that’s currently the active design frontier.