Katia Gil Guzman is a developer-experience engineer at OpenAI and former co-founder and chief technology officer of Emera, a French software startup. She helps developers put AI coding agents to work across repositories, business applications, visual development, and recurring operational tasks.
From Microsoft to startup founder to OpenAI
After studying at Télécom SudParis, Gil Guzman joined Microsoft France, first as an intern and then as a software engineer in its Commercial Software Engineering organization. She worked alongside customers’ development teams on Azure-based solutions involving conversational systems, natural-language processing, document understanding, and computer vision. A 2019 interview captures her early commitment to making AI useful outside research environments.
In April 2021, she left Microsoft to co-found Emera as CTO. She wrote about the transition from corporate engineering to startup leadership, where product judgment, customer discovery, and technical execution became inseparable. She subsequently worked at Stripe as a solutions architect before joining OpenAI in a solutions-architecture role and moving into developer experience.
Making coding agents useful in practice
At AI Engineer Europe 2026, Gil Guzman demonstrated several practical principles for working with Codex:
Reusable agent workflows: Plugins package skills, connected applications, and MCP servers so developers can repeat established processes without reconstructing their tools and instructions each time.
Coding agents beyond code generation: Effective software agents need repository awareness, command execution, usable interfaces, environmental safeguards, and connections to services such as Slack and GitHub.
Scheduled operational automations: Agents can surface urgent Slack conversations and emails or transfer event information from a website’s source files into a Google Drive spreadsheet.
Visual development with browser-based validation: Combining image generation with interactive Playwright sessions lets an agent create game assets, inspect the running interface, and test its behavior.
She also highlighted a consequential security distinction: an agent may be able to inspect resources committed to a repository without being authorized to execute the scripts packaged inside a skill.
A practical walkthrough of Codex’s shared harness, plugins, automations, code review and subagents, following live demonstrations through their results and limits.
What does an engineering agent need beyond code generation?
