George Millo on Agentic Coding

14 snips

Mar 22, 2026

George Millo, creator of Learn Phoenix Live View and Gauntlet AI alumnus, is an Elixir educator and engineer focused on AI-first workflows. He discusses agentic coding realities like verification debt, planning before prompting, and working in small self-contained PRs. He explores Elixir and the BEAM as natural fits for agent architectures and warns about security risks when giving AI tools codebase access.

Ask episode

AI Snips

Chapters

Transcript

Episode notes

ADVICE

Plan Before You Prompt

Plan thoroughly before prompting so you and the LLM share a clear mental model of the feature.
George outlines architecture, edge cases, and file locations with Codex/Claude before asking the model to implement changes.

INSIGHT

Verification Debt Is The New Risk

Verification debt grows when your mental model lags behind AI-generated code.
George warns that cognitive gaps lead to unsafe or incorrect systems if you let the AI run ahead without checks.

ADVICE

Use Tests As Safety Rails Not Checkboxes

Use tests and constraints as rails to prevent the AI from hacking around requirements.
George cautions against naive test-first prompts because LLMs may craft superficial tests unless you verify their substance.

Get the Snipd Podcast app to discover more snips from this episode

Get the app

George Millo — creator of Learn Phoenix Live View and alumnus of the Gauntlet AI fellowship — joins me to talk through what really changes when you commit to LLM-driven development. George went through Gauntlet's intensive 10-week program built around LLM maximalism, came out the other side rethinking how he builds software, and has spent the past year working on an AI-first engineering team applying those lessons in production.

We spend a lot of time on the practical realities of agentic coding: verification debt (the gap that grows between your mental model and what the AI actually built), the importance of planning before prompting, why George works in small self-contained PRs, and how he uses Codex and Claude Code in parallel tabs without losing track of what's happening. We also get into the debate around vibe coding, spec-driven development, testing pitfalls, and why deep technical knowledge matters more now — not less — when AI is writing most of the code.

The conversation covers where Elixir fits in an AI-first world: the BEAM's process model as a natural fit for agent architectures, Phoenix shipping with an agent.md file, Tidewave's approach to closing the feedback loop, and why Elixir's tooling consistency puts it ahead of the JavaScript fragmentation George deals with at his day job. We also get into the security risks that come with AI-assisted development — giving LLM tools codebase access, the OpenClaw skills marketplace vulnerabilities, and the kinds of security mistakes that are now much easier to ship without noticing.

George closes with practical advice for anyone hesitant to adopt agentic workflows: stay curious, ask the AI to explain the code it writes, build something outside your comfort zone, and put in the reps.

Resources Mentioned:
- Learn Phoenix Live View: https://learnphoenixliveview.com

Connect with George:
- X/Twitter: x.com/georgemillo
- LinkedIn: linkedin.com/in/millog

Sponsors:
- Paraxial.io — Elixir-first application security: paraxial.io
- Jido — Elixir AI Collective Discord: agentjido.xyz/discord

SUPPORT ELIXIR MENTOR
- Elixir Mentor: elixirmentor.com