Shippy tests the agent harness

Ai2's Shippy writeup describes a research assistant built around deterministic tools, source links, timestamps, deep links, hosted isolation, and agent-level evaluation. Why it matters: The model is only one part of a useful research agent. The harness decides what evidence is shown, where the run is isolated, and whether the output can be checked after the chat ends.

Try this: Use the post as a review checklist: deterministic tools, cited sources, isolated execution, deep links, and evals that test the full agent loop.