Field guide / Practical workflows

Dogbench

Experimental local tooling for agent usage comparisons. No public scoreboard, guaranteed savings, or model-quality ranking.

Dogbench collects usage from supported agent clients and explores token, timing and cost comparisons. Check the current source and setup requirements before running it; provider-backed runs can incur usage charges.

What a useful report needs

• Pinned task, inputs, model/provider and runtime.
• Clear cache state, retries and included tool/output traffic.
• Actual usage distinguished from estimates and missing data.
• Task correctness, repeated runs, variance and sanitized evidence.

The website has not independently reproduced the complete Dogbench provider/comparison surface. Older README sample numbers are not live results.