Demand evidence from research sub-agents
Technique
When an agent hands research to sub-agents (helper agents it launches), tell them to read the actual code behind every claim and never cite a file they didn't open. They should also not edit anything. Muse does this on its own.
Theo's current stance
Theo calls it a really good instruction to give a sub-agent. It's part of why he rates Muse as a strong orchestrator, which labs other than Anthropic and OpenAI usually struggle with.
Try it
- Ask your agent to tell research sub-agents to open and read every file they cite.
- Make research sub-agents read-only so they report findings without changing files.
First seen
16:42
Requirements, read actual implementation bodies for every claim you site. Search results only locate candidates. Do not cite a file you did not open. Map, yada yada yada. Do not edit any files as announced only. This is a really good thing for a model to send to a sub-agent, which makes this model really
Every mention
16:42
Requirements, read actual implementation bodies for every claim you site. Search results only locate candidates. Do not cite a file you did not open. Map, yada yada yada. Do not edit any files as announced only. This is a really good thing for a model to send to a sub-agent, which makes this model really
Theo Daily is unofficial and not affiliated with Theo or t3.gg. Written by Claude Opus 5.5 from each video's transcript.