LLMs Can't Write Great Software, So We Verify Everything - Ohans Emmanuel, Coldtea
Ohans Emmanuel spends 93% of his coding time with agents, and he has reached a somewhat sad conclusion: LLMs fundamentally cannot write great software. At the TypeScript AI Conference in London, the Coldtea co-founder shows how his team still gets decent software out of them. His answer is a zero-trust harness. You trust nothing the agent hands you and verify it at every stage of the development lifecycle. How does it work? Planning produces an implementation plan that sits on the task, where the team reviews it before any code exists. Review is split between models, so whatever Claude builds, Codex reviews, and the other way round. Visual QA agents open a real browser or mobile device on every pull request and click through the app like a user would. Background agents watch production signals from tools like Sentry and PostHog and open tasks with the fix already investigated. Recorded at the TypeScript AI Conference in London, July 23, 2026, presented by Mastra. 🔗 MENTIONED IN THIS TALK Coldtea: https://www.coldtea.ai Coldtea docs: https://www.coldtea.ai/docs 📚 MASTRA RESOURCES https://mastra.ai https://x.com/mastra_ai https://mastra.ai/community/discord https://github.com/mastra-ai Learn Mastra: https://mastra.ai/learn https://mastra.ai/books/principles-of-building-ai-agents https://mastra.ai/books/patterns-of-building-ai-agents WHAT IS MASTRA Mastra is an open-source TypeScript framework designed for building and shipping AI-powered applications and agents with minimal friction. It supports the full lifecycle of agent development — from prototype to production. CHAPTERS 00:00 Intro 00:23 LLMs can't write great software 01:22 Between giving up and YOLO 02:11 Ohans Emmanuel and Coldtea 02:36 Don't trust, verify 03:07 Building was never the hard part 03:46 Planning 04:19 Reviewing plans before pull requests 06:20 One model builds, another reviews 06:58 The review loop 08:26 Visual QA agents on every pull request 10:50 Exploratory QA 11:43 Post-deployment and background agents 13:09 Guardrails at every checkpoint