Grok Bot 工程师分享 6 个 agent 技巧

Diam @diamai_原推发布于 2026.10.17

在 X 上打开原推

The engineer building Grok Bot at SpaceX gave Stanford the full playbook in 45 minutes. I pulled the 6 tricks you can steal for Claude Code or Codex tonight. What his agent does that yours probably doesn't: → Only tool names go into each message, the full schema sits in a file until the model needs it → A cache miss counts as a bug. System prompt and tools stay frozen, sections keep a fixed order → Walk away from a 100K-token chat and it compacts right then, while the cache is still warm and cheap → Screenshots and long shell sessions go to subagents, the main agent only gets the result → Before it bothers you, it tries memory, plugins, web search, then a browser. You're the last step → If the model spends 90% of its time on one kind of shell command, give it a dedicated tool Lee Robinson, who gave the talk, says that in a blind test, a model reviewing every terminal action would beat the humans approving them most of the time. Paste this into your CLAUDE.md or AGENTS.md: - Write any output longer than 50 lines to a file and give me the path - Send screenshots, log reading and multi-step shell work to a subagent, bring back only the result - Before asking me, check memory, then MCP tools, then web search - On web pages, read the accessibility tree before taking a screenshot Full lecture below, the harness part starts at 18:07. Which of these is your agent still missing?

这条是引用着另一条写的：x.com/diamai_/status/2111105993341882565