
Your AI agent looped all night. Blame the harness. #shorts #aiagents #llm
Your AI agent loops, burns through tokens, hits the quota, and then tells a customer something that was never true. Nobody saw it happen. An AI agent is a model plus a harness: the tools, context, prompts, memory and guardrails around it. When an agent misbehaves, the harness is usually the culprit. OpenLIT is the open source agent harness engineering platform: trace it, evaluate it, fix it, verify it. In this video: 0:00 The crime scene: a looping agent, a 429 quota error, a wrong answer 0:14 What developers keep posting about 0:22 Enter the detective 0:29 Observe the harness: docker compose up -d, pip install openlit, openlit.init() 0:40 Clue 1: the loop (1.9M tokens for one task) 0:48 Clue 2: generation issues on the trace (truncated, filtered, model swapped) 0:55 Clue 3: cost spike and token usage spike alerts 1:03 Case closed: harness fixed, same task 79% cheaper The before and after numbers are two real Claude Code runs of the same task, measured in OpenLIT. The harness fixes were project context, model routing, and a scoped prompt. The opening terminal scene is a reconstruction. Get started: https://github.com/openlit/openlit Docs: https://docs.openlit.io Website: https://openlit.io What is your agent hiding? #Shorts #AIAgents #AgentHarness #OpenTelemetry #OpenSource












