Skip to content

Understanding Outputs

After a run completes, results are saved to outputs/<agent_id>/.

Output Structure

outputs/<agent_id>/
├── best_iteration_snapshot/           # Best iteration snapshot
│   ├── model_training/
│   │   ├── train.py          # Training script
│   │   └── training_artifacts/
│   ├── model_inference/
│   │   └── inference.py      # Inference script
│   ├── validation_evaluation/
│   │   ├── eval_predictions_train.csv
│   │   ├── eval_predictions_validation.csv
│   │   └── output.json
│   └── runtime_info/
│       ├── iteration_metadata.json
│       ├── environment.yml
│       └── environment.tar.gz         # Present in full export mode
├── run/                      # All iterations + shared run state
│   ├── shared/
│   │   ├── config.json
│   │   ├── dataset_metadata.json
│   │   └── splits/
│   ├── iteration_0/
│   ├── iteration_1/
│   └── ...
├── reports/
│   ├── best_iteration.md
│   ├── best_iteration.pdf
│   ├── markdown/
│   │   ├── run_report_iter_0.md
│   │   ├── run_report_iter_1.md
│   │   └── ...
│   └── pdf/
│       ├── iteration_0.pdf
│       ├── iteration_1.pdf
│       └── plots/
├── logs/                     # Logs and metrics
└── README.md                 # Run summary

best_iteration_snapshot

The most important directory - contains the best-performing iteration's artifacts.

File Description
model_inference/inference.py Script to run predictions
model_training/train.py Script that trained the model
model_training/training_artifacts/ Trained model files (format varies)
runtime_info/iteration_metadata.json Which iteration produced the snapshot
runtime_info/environment.yml Portable definition of the Conda environment used
runtime_info/environment.tar.gz Packed environment for fast container-local restoration in full mode
eval_predictions_test.csv Best-model predictions for the optional test split
eval_predictions_test.numeric_labels.csv Numeric test labels used for metrics
eval_predictions_test.metrics.json Metrics for the optional test split

Multiple held-out test sets are also supported. Each uses its directory name in place of test in the filenames above. See Evaluating on Held-out Test Sets.

Using the Best Model

agentomics-inference --agent-dir outputs/<agent_id> --input data/input --output predictions.csv

Iteration Directories

run/shared/dataset_metadata.json stores the resolved dataset preparation metadata. Forks reuse it so task type, label mapping, and the CSV label column are not requested again.

Each iteration's files are preserved under run/iteration_N/:

run/iteration_N/
├── model_training/
│   ├── train.py
│   └── training_artifacts/
├── model_inference/
│   └── inference.py
├── runtime_info/
│   ├── environment.yml
│   ├── iteration_metadata.json
│   └── iteration_state.json
└── ...                       # Other iteration artifacts

Reports

reports/best_iteration.md and reports/best_iteration.pdf are copies of the selected iteration's reports and provide stable paths to the final result.

Iteration Reports

reports/markdown/run_report_iter_N.md - Summary of each iteration:

  • Data exploration findings
  • Model architecture chosen
  • Training details
  • Validation metrics

PDF Reports

reports/pdf/iteration_N.pdf - PDF report per iteration, including metrics and plots for every evaluated held-out test split, plus plots in reports/pdf/plots/.

Metrics

Metrics are tracked for each iteration:

Metrics depend on the selected validation metric and task type. See agentomics-run --list-metrics for the current list.

Where Outputs Are Stored

The agent writes directly to the run workspace as it works — there is no separate staging area or temporary volume:

  • The host workspace defaults to outputs/<agent_id>/ and is mounted at /workspace in the container.
  • Dataset preparation directories may exist there while the run is active, but they are removed after test evaluation and reporting. Exact versioned splits remain under run/shared/splits/.

W&B Logging

If W&B is configured, you'll also find:

  • Experiment tracking at wandb.ai
  • Agent traces with Weave
  • Metric plots and comparisons
  • Artifact versioning

See Environment Variables for W&B setup.

Reproducing Results

To reproduce a run:

  1. Use the same dataset
  2. Use the same model and parameters
  3. Set the same random seed (if applicable)

The model_training/train.py and model_inference/inference.py scripts contain all logic needed to reproduce the model.

Cleaning Up

Remove old runs:

# Remove specific run
rm -rf outputs/<agent_id>

# Remove all runs (careful!)
rm -rf outputs/*

Next Steps