Our summary
An AI assistant often needs to gather information before making a judgment. Earendil explains why Pi now supports MCP, a way to connect tools, and Codemode, a small JavaScript workspace for combining their calls. Its example asks Jev to find frustration in issue-tracker comments.
Connected tools fetch open issues and their comments from Linear. Jev chooses between neutral, mildly frustrated and highly frustrated, with instructions to ignore how serious the bug is. Four workers handle separate issues at once. Ordinary code saves the answers and sorts flagged threads using their probabilities.
The article shows a condensed replay covering 167 issues, not a measured accuracy test. Access to Linear and a provider offering Jev is required. The example illustrates how to keep retrieval, judgment and sorting separate; it does not establish that the tone labels are correct.
Key takeaways
- Let connected tools gather the information; give Jev a narrow question and clearly described answer options.
- Separate emotional tone from the seriousness of a bug so the model judges the intended property.
- Run independent work in small concurrent groups and save the results before sorting them in ordinary code.