Skip to main content
AgentCompass writes each run under the configured results_dir, defaulting to results.

​
Details

details/ contains per-task JSON files. Completed samples use the task id as the filename. Retryable failures may be written as _error_<task_id>.json so reuse mode can retry them later. If analysis is enabled, each attempt can include:

​
Summary

summary.md contains aggregate benchmark metrics, run metadata, and high-level status. Regenerate it without running analyzers:
Use --dry-run to print the regenerated markdown without writing files.

​
Analysis Summary

analysis_summary.json and analysis_summary.md aggregate analyzer output across tasks. Recompute them with:

​
Resume and Reuse

When reuse is enabled, AgentCompass skips completed details with matching effective parameters. Execution-only knobs such as task concurrency do not need to force a new result directory.

​
Progress

progress.json records the latest structured state. progress.jsonl records the event stream and is the best source for external monitors.

​
Local Result Browser

For a local web view of summary.md, progress, and per-task details, start the result browser from the repository checkout:
Open the Vite URL printed by the dev server and enter the absolute run directory path, such as /path/to/AgentCompass/results/<benchmark>/<model>/<run>. The path is resolved on the machine running npm run dev, including when the UI is reached through port forwarding.