canwhite/AgentEval

What it solves

AgentEval provides a transparent way to monitor, evaluate, and debug AI agents. It solves the difficulty of understanding why an agent failed a task by capturing the traffic between the agent and the LLM, automatically organizing it into sessions, and providing automated grading and root-cause analysis.

How it works

It operates as an HTTP proxy that sits between your AI agent and the upstream LLM API. It captures all raw traffic and processes it through several stages:

1. Session Detection: It automatically splits traffic into distinct sessions based on idle timeouts or when it detects a new conversation starting.
2. Grading: It uses a combination of rule-based metrics and an LLM judge to score sessions across four dimensions: task completion, tool efficiency, response quality, and performance.
3. Diagnosis: A rule engine runs 10 specific checks to detect behavioral issues like tool chain breaks, retry loops, or token waste.
4. Probing: An LLM agent with read-only access to the agent's source code directory reviews configuration files (prompts, skills, tools) to identify the root cause of the diagnosed issues and suggest improvements.
5. Visualization: All results are presented in a built-in web dashboard for easy review.

Who it’s for

Developers building AI agents who need a systematic way to evaluate performance, diagnose behavioral bugs, and optimize their agent's prompts and tool configurations.

Highlights

• Transparent Proxy: No need to modify agent code; just change the BASE_URL.
• Multi-dimensional Scoring: Combines hard rules with LLM judgment for a comprehensive 0–1 score.
• Automated Root-Cause Analysis: The "Probe" feature analyzes source code to explain why a failure occurred.
• Behavioral Diagnosis: Detects specific failure patterns (e.g., silent failures, duplicate tool calls) using a dedicated rule engine.
• Integrated Dashboard: Provides a structured view of conversations, grades, and diagnosis reports.