You Should Write An Agent (fly.io)

A developer argues you should just build an LLM agent to really understand them — and proves it by walking through a tiny, working example using the OpenAI Responses API (model "gpt-5"). The core idea is simple: keep a context as a list of role-tagged messages, call the stateless LLM endpoint in a loop, and let the model return special "function_call" outputs that your loop executes (e.g., a ping tool). The article shows multi-personality behavior by switching system prompts, demonstrates wiring tools via a JSON function schema, and explains a small handler that resolves function calls by running the native tool, returning a function_call_output for the model to consume. From REPL chat to ping/traceroute or bash-running coding agents, the author shows you can get a capable agent up and running in minutes. The piece’s significance is twofold for the AI/ML community: first, agents are surprisingly easy and extremely informative — building one teaches practical limits (stateless models, token budgets) and design patterns (sub-agents as separate context arrays, summarization/compression, tool-specific contexts). Second, it reframes debates about plugin ecosystems (MCP) and security: plugins mainly save a bit of glue code but cede control; rolling your own agent gives finer-grained context, tool-scoping, and security architecture. Practical implications include rapid prototyping (coding agents, vulnerability scanners, orchestration), the need for context engineering to manage token usage, and rich opportunities for experimentation and new startups.