Topic: What to Hand to an Agent and What Humans Keep, Seen Through My Own Incident Logs
The line on the cover, PASS T03 all ArgoCD apps Synced+Healthy, was a pass, yet there were actually zero ArgoCD apps. Following cases of this kind of silent failure, the talk looked at where a developer's work moves in the age of AI agents. The slides are in Korean.
Learning goals
- Explain how an AI agent works (choosing and calling tools, context and harness, multi-agent setups).
- Describe an agent development environment (model, context, harness) at the level of its configuration.
- Recognize how agents fail silently, and verify a "done" report against evidence.
- Explain the division of roles between humans and AI (what to hand off and what to keep, responsibility versus prevention) and apply it to your own project.
Outline
- My AI journey: how my way of using AI changed, from ChatGPT to Claude Code, Antigravity CLI, and Codex CLI
- What an agent is: how it differs from a chatbot, models, context and harness, the Ralph loop, hub-and-spoke multi-agent setups
- How agents fail silently: cases 1 to 4 and a feedback structure (incident, lesson, discriminating criterion, regression test)
- Dividing roles between humans and AI: an open-source contribution case, what to hand off and what to keep, responsibility and prevention
Key messages
- The person who approves takes the responsibility, and a mechanism prevents the recurrence.
- Put a floor on the count in "all are fine" checks, treat a silent SKIP as a FAIL, and require execution evidence in every completion report.
- The cases were picked from my own records, so they are not generalized into a property of AI.