EvalForge

The Problem

Testing agentic AI systems purely by asserting final string responses masks severe failure modes: tool misrouting, hallucinated parameters, unfaithful grounding, and explosive latency or token consumption across multi-turn decision loops.

Key Decisions

• Full-Trajectory Observability: Captured granular intermediate execution traces (thoughts, tool invocations, arguments, observations, token usage, latency, and cost).
• Dual-Pronged Evaluation Subsystem: Paired deterministic heuristics (token budgets, latency ceilings, tool schema validation) with cognitive LLM-as-a-judge rubrics for faithfulness and groundedness.
• Experiment & Regression Tracking: Built automated comparative regression delta analysis against version-controlled golden datasets and experiment baselines.

Results

• Observability: End-to-end trace waterfalls and per-step cost/latency breakdowns.
• Repeatability: Quantitative regression detection preventing silent degradation during prompt or model upgrades.

What I Learned

Agent evaluation must transition from subjective manual inspection (“vibes-based testing”) to a quantitative, repeatable engineering discipline backed by versioned benchmarks.

[Built as an agent evaluation platform. Source code is public.]