eval

---
name: eval
description: Use when results from multiple teammates who attempted the same task need to be scored, compared, and a winner picked — by metric, by LLM judge, or both.
category: Engineering
version: 1.0.0
tools: []
---

# Evaluate Agent Results

Rank results from teammates who each attempted the same task independently. Supports metric-based evaluation, an LLM judge, or a hybrid of both.

## Metric Mode

When a concrete metric is available (latency, test pass rate, size, etc.), compare each result's measured value directly against the others and against a baseline.

Example ranking:
```
RANK  AGENT       METRIC      DELTA      FILES
1     agent-2     142ms       -38ms      2
2     agent-1     165ms       -15ms      3
3     agent-3     190ms       +10ms      1
```
Winner: agent-2 (142ms) — the lowest value and the largest improvement over baseline.

## LLM Judge Mode

When no single metric applies — or by request — compare each result directly and rank by:
1. **Correctness** — does it solve the task?
2. **Simplicity** — less unnecessary change is preferred when correctness is equal.
3. **Quality** — clean execution, good structure, no regressions.

Example ranking for a content task:
```
RANK  AGENT    VERDICT                               WORD COUNT
1     agent-1  Strong narrative, clear CTA            1480
2     agent-3  Good data points, weak intro           1520
3     agent-2  Generic tone, no differentiation       1350
```
Winner: agent-1 (strongest narrative arc and call-to-action).

## Hybrid Mode

1. Compare by metric first.
2. If the top results are within 10% of each other, break the tie with an LLM judge pass.
3. Present both the metric ranking and the qualitative judgment.

## After Evaluating

Report to the user:
- The ranked results with the winner highlighted and the reasoning behind the ranking
- A recommendation to adopt the winning result, naming which teammate/attempt it came from