OpenAI AgentKit Explained Why Agent Builders Fail in Production | Dev in the Details

Building AI agents has never been easier.Shipping them safely to production has never been harder. In this episode of Dev in the Details, I sit down with Predibase co-founder and CTO Travis Addair to break down what OpenAI’s AgentKit launch actually means for teams trying to deploy agents in the real world. We unpack: * What OpenAI announced at Developer Day — and what AgentKit really is * Why agent builders, drag-and-drop UIs, and “5-minute agents” break down in production * How AgentKit compares to RPA tools (UiPath), automation platforms (Zapier), and open-source frameworks like n8n * The difference between no-code, low-code, and fully custom agent systems * Why evals, guardrails, CI/CD, and governance are the real blockers to production * How agent evaluation is different from LLM evals — and why “LLM as a judge” may actually work better for agents * What’s coming next: reinforcement learning and continuous agent improvement If you’re a CTO, CIO, AI engineer, platform lead, or enterprise builder, this episode explains why most agent demos stall — and what it actually takes to run agents at scale. 🎧 Subscribe for future episodes on agent governance, evals, and reinforcement fine-tuning. 00:00 – Why AI agents are easy to build but hard to ship 01:05 – OpenAI Developer Day recap & AgentKit overview 02:10 – What AgentKit includes: canvas, connectors, evals 03:20 – Are AI agents really new? RPA, Zapier, and history 05:10 – Who does OpenAI AgentKit compete with? 06:40 – AgentKit vs UiPath, Zapier, and n8n 08:35 – Why AgentKit feels more like a chatbot builder 10:10 – Why RPA and agent platforms are on a collision course 12:00 – Build vs buy for AI agents in the enterprise 14:20 – Why no-code ML tools historically failed 16:00 – Prediction: no-code vs low-code vs custom agents 18:40 – Where real enterprise value from agents is created 21:00 – What agent builders usually ignore 23:30 – Why production agents are still “just software” 26:10 – CI/CD, versioning, and trus