Descriptions:
Kieran Klaassen, CTO of Kora at Every, joins Peter Yang to demonstrate compound engineering — a systematic methodology for making Claude Code progressively smarter across sessions rather than treating each interaction as a clean slate. The core insight is that every Claude Code session produces learnings that, if codified into the project’s CLAUDE.md file, directly shape the agent’s behavior in all future sessions, creating a compounding improvement loop.
The four-phase cycle — plan, work, assess, codify — is implemented through a custom plugin Klaassen built and made publicly available. The plugin automates the assessment and codification steps, including a single `LFG` command that runs tests, generates documentation, and opens a pull request as one autonomous action. Klaassen demos the workflow on Kora, his AI email assistant, targeting a specific feature: making the in-app agent capable of performing any action a human user can, a concept he calls being “agent-native.”
A recurring theme is the shift in trust dynamics between engineers and AI-generated code. Klaassen argues that experienced AI-native engineers build reliable systems and documentation structures instead of reviewing every line, and that CLAUDE.md functions as persistent institutional memory — auto-updated by Claude itself when instructed to “add this to CLAUDE.md.” The conversation also covers managing multiple parallel Claude Code agents, the tradeoffs of bypass-permissions mode versus manual approval, and how to structure context for agents working on production codebases with real user traffic.
📺 Source: Peter Yang · Published February 08, 2026
🏷️ Format: Workflow Case Study
