Loading…
Friday August 14, 2026 11:10 - 11:35 KST
Using AI efficiently begins with the outcome—but model choice is only one part of the design decision. As coding agents become more capable, teams must also decide when an agent should explore and generate code, when a validated method should become a reusable template, when a fixed application only needs an API, and when multiple agent experiences justify a shared MCP interface.

Drawing on recurring questions from the Codex developer community, this session connects two examples. An early-stage Science Agent Leaderboard project from the PseudoLab community illustrates how agents can build executable solutions, and why generated work must be evaluated for correctness, cost, and reproducibility rather than judged by the answer alone. A public, synthetic rare-defect analytics example then shows how a data-science method can move from Codex-assisted exploration to evaluation, human review, versioned code, and reusable templates.

The session concludes with an illustrative air-gapped architecture in which fixed analytics applications use APIs, while on-prem coding agents and analytics chatbots can access selected, validated capabilities through a small MCP surface.
Speakers
avatar for Junho Kong

Junho Kong

AI Platform Architect, SK On
Junho is an  at SK On and a Codex Ambassador
working at the intersection of agentic engineering, AI platforms, developer tools, and community education. He leads and supports developer community initiatives around Codex and AI-assisted software development, helping developers move from ad-hoc AI usage toward more structured... Read More →
Friday August 14, 2026 11:10 - 11:35 KST
Grand Ballroom 2 + 3

Sign up or log in to save this to your schedule, view media, leave feedback and see who's attending!

Share Modal

Share this link via

Or copy link