The Three-Layer Agent Stack
by Alexander Embiricos • Product Lead for Codex at OpenAI
Former startup founder (screen sharing/pair programming), PM at Dropbox
🎙️ Episode Context
Alexander Embiricos explores the evolution of Codex from a code completion tool to a proactive 'Software Engineering Teammate'. He discusses the internal strategies OpenAI uses to build agents (The Three-Layer Agent Stack), the shift in developer workflows towards 'Vibe Coding' and 'Chatter-Driven Development', and shares case studies of massive acceleration, such as building the Sora Android app in under a month.
Problem It Solves
Addresses the disconnect between raw model capabilities and actual product utility, ensuring models are usable, context-aware, and persistent.
Framework Overview
A framework for building effective AI agents by synchronizing innovation across three distinct layers: the model intelligence, the API interface, and the product harness. Success requires tight integration rather than treating the model as a black box.
🔺 Priority Pyramid
When to Use
When building AI-powered products or agents where raw intelligence is not enough to solve user problems.
Common Mistakes
Optimizing only the model without adapting the harness/UI; assuming a generic API will support complex agentic behaviors like long-running tasks.
Real World Example
Implementing 'Compaction' to allow Codex to run for 24 hours required the Model to understand summarization, the API to handle the handoff, and the Harness to prepare the payload.
It turns out lets you just do a lot more and try many more experiments as to how these things will work together... shipping this compaction feature... actually meant working across all three things.
— Alexander Embiricos