I Built a Web Search Agent Harness. Then I Checked If It Actually Deserved the Name.

Short summary

A developer builds a Perplexity-style web search agent using Bun, React 19, Tavily, and OpenRouter, then critically audits whether the codebase deserves to be called an agent harness. The post defines a harness as scaffolding that lets the model decide tool usage, loops until completion, trims context, and outputs structured events — and concludes the project qualifies, albeit as a single-tool harness with honest caveats about its limitations.

• •Built a single-tool web search agent with streaming answers and clickable citations
• •Defines what makes code a harness vs. a script: model decides tool use, loop iterates, context is trimmed, output is structured
• •Honest about limitations: one tool, one provider, no retries or fallback yet

Generated with AI, which can make mistakes.