Modes
Fast for triage, Deep Research for root-cause work. Tool permissions control what runs without approval.
Two controls: response mode (how hard the agent thinks) and tool permission mode (what it can do without approval).
Fast vs. Deep Research
Deep Research is the default. It plans a multi-step investigation: searches across traces, scores examples, cross-references behaviors, and stitches evidence together before answering. Right for:
- Root-cause analysis across many traces
- Judge prompt iteration grounded in real scoring
- Test run comparisons with example-level diffs
- "Why is detection rate low?" style questions
Deep Research takes longer and returns a more thorough answer with multiple citations.
Fast is the lighter alternative. Single-pass answers grounded in the current page snapshot plus a small number of tool calls. Right for triage, summaries, and "what's on this page" questions.
Tool permissions
The agent has real tools: search traces, score examples, draft rubric and behavior changes. The tool permission mode controls whether write tools require approval.
- Ask for writes (default). Confirm before tools create, update, or delete data.
- Auto-allow writes. Write tools run without an approval card.
Read-only tools always run without prompting in either mode.
Next steps
- Use Cases. Concrete flows paired with the right mode and permissions.
- Context & Mentions. What the agent already sees before any setting changes.
Context & Mentions
Page context auto-attaches to every Judgment Agent message. Use @ to pin specific judges, behaviors, and automations.
Use Cases
Common questions Judgment Agent is best at answering. Top-down investigations, trace debugging, judge iteration, behavior tuning, eval comparison.
Last updated on