@TiagerBao posts Chinese notes on a third-party compilation, "Jev Engineering for Coding Agents," of TypeSafe founder Diogo Almeida's thinking. He writes that writing code is only about 4–10% of tokens, and that the notes are not a shipped complete plan.
- 11.0k views
- 168 likes
- 189 bookmarks
- 20 reposts
About this case
Guide. 旺旺 (@TiagerBao) says more tokens go to reading, search and command output. Points he flags: a small model may not save money, because rebuilding context after a switch can cost more; MCP tools should not load every schema at once; compaction should extract context for the current task instead of compressing blindly; AGENTS.md should load conditionally so unrelated rules do not sit in context; sub-agents should share retrieval results. The core idea he names is that context should be assembled for the task, not left to accumulate as chat history. He says the PDF is third-party design notes worth reading for people who build agents, not a complete plan already shipped. The original PDF is https://github.com/yibie/jev-engineering-zh/blob/main/assets/original.pdf. He asks whether the next gain is model ability, or context management and task scheduling. Hashtags on the post are #Agent, #ClaudeCode, #Codex and #AI. Text only.
Original posthttps://x.com/TiagerBao/status/2120041783568875653