# Compare two AI agent configurations

Blank worksheet. No measured results are supplied. A configuration includes the model, runtime, tools and settings.

## Define the trial before running either configuration

- Decision this trial should inform:
- Task and reason for choosing it:
- Project copy / dataset / starting version:
- Reproducible defect or starting condition:
- Expected result:
- Acceptance criteria:
- Edge case and required evidence:
- Allowed tools, network and actions:
- Time and spending limit chosen for this trial:
- Identical instructions given to both configurations:
- Reviewer and review method:

## Configuration, for each run

- Run identifier:
- Date and time zone:
- Provider / exact model / available version:
- Runtime and version:
- Reasoning, context and tool settings:
- Hardware and quantisation for a local model:
- Environment differences between candidates:

## Observation, for each run

- Start and finish:
- Produced result and link:
- Accepted, failed and unverified criteria:
- Evidence inspected by the reviewer:
- Follow-up prompts and human interventions, with reasons:
- Observed review and correction time:
- Actual billed cost / observed usage / unknown:
- Failure, interruption or limitation:

## Decision

- Verdict for this task: accepted / rejected / unverified
- Keep the configuration, retry or restrict it to a role:
- Reasons and limits of the conclusion:
- Changes that would justify a new trial:

Keep failed runs in the CSV log. Do not substitute zero for missing measurements. Do not infer a general ranking from one task. If tools or runtimes differ, compare complete configurations rather than models alone.

Guide: https://stellary.co/blog/best-ai-coding-models/