judge-trajectory

Evaluate agent trajectories and tool-use traces with process-aware rubrics.

Updated Mar 25, 2026
One-click install
npx skills add https://github.com/Tyler-R-Kendrick/copilot-auto-training --skill judge-trajectory-tyler-r-kendrick
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: judge-trajectory
Source: https://github.com/Tyler-R-Kendrick/copilot-auto-training/tree/main/skills/judge-trajectory
Command: npx skills add https://github.com/Tyler-R-Kendrick/copilot-auto-training --skill judge-trajectory-tyler-r-kendrick

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) and assets (resource) components.

What problem does it solve? Judging agent runs by final answers alone misses broken tool calls, weak evidence gathering, and runtime failures, leading to verdicts that reward polished output over reliable process. ## Core Features & Use Cases - Process-Aware Rubrics: Lock 4-7 trajectory dimensions covering plan suitability, evidence gathering, tool correctness, failure handling, side effects, and final outcome quality. - Evidence Ledger Workflow: Build trace evidence from tool calls, intermediate artifacts, and failure logs, then score candidates against one stable rubric with order-robustness checks. - Use Case: Compare two code-fix agent runs where one produced a correct patch through repeated broken tool calls, and surface how runtime failures change the verdict. ## Quick Start Use the judge-trajectory skill to compare these two agent traces and decide which run is stronger based on process evidence, not just the final answer.

Frequently Asked Questions about judge-trajectory

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I evaluate agent trajectories instead of just final answers?

Lock a process-aware rubric with 4 to 7 dimensions such as plan suitability, tool correctness, and failure handling, then build an evidence ledger from tool calls and intermediate artifacts. Score every candidate against the same rubric so process quality becomes first-class evidence.

How to compare two agent runs with tool traces?

Gather both trajectories including tool calls, failure logs, and intermediate artifacts, then score them against one locked rubric. Separate operational failures from quality failures and run an order-robustness check before finalizing the verdict.

When should I use trajectory judging instead of outcome-only judging?

Use trajectory judging when process reliability matters, such as agent runs with tool calls, planning, or side effects. For clean outcome-only response comparison where the final answer is the only signal, switch to an outcome-focused judging contract instead.

Should I trust chain-of-thought reasoning when judging agent runs?

Treat narrated chain-of-thought as low-trust evidence unless trace data corroborates it. Prefer observable artifacts, tool call records, and verifier-backed signals over self-reported explanations when scoring trajectories.

What are the limitations of trajectory-based evaluation?

Thin traces or incomplete artifacts limit confidence, so verdicts should preserve uncertainty rather than overclaim. Process judgments can also be biased by presentation order, which is why an order-robustness check is required before finalizing.