Running the Agent¶
The agent is launched through agentomics-run, which starts the Agentomics
Docker image and mounts the required datasets and workspace. See
Installation for setup.
Interactive Mode¶
With no run arguments (and an attached terminal), the agent prompts you for the essentials:
You'll be prompted to select:
- LLM Model - Choose from available models
- Dataset - Select a dataset
- Iterations - Number of optimization cycles (default prompt: 5)
The validation metric is not prompted interactively; pass --val-metric to
override the task-based default (AUROC for classification, MAE for
regression).
Non-Interactive Mode¶
Supply parameters directly to skip the prompts:
For non-interactive fresh runs, provide at least --model and --dataset. If
you omit --iterations, the default is 5.
Options¶
agentomics-run exposes many flags — model and provider selection, time limits, split
and exploration controls, forking, and more. See CLI Options
for the complete reference, or run:
A few common ones:
# Set a deadline for the whole run (whichever of this or --iterations hits first)
agentomics-run --timeout 7200 --model openai/gpt-5.1-codex-max --dataset my_data
# Control how long the agent explores baselines / may re-split the data
agentomics-run --split-allowed-iterations 1 --exploration-iterations 4 ...
# Override the optimization goal
agentomics-run --user-prompt "Only use simple models like logistic regression" ...
See Custom Prompts for prompt tuning and Forking a Run for branching off an existing run's checkpoint.
What Happens During a Run¶
- Dataset Preparation - Validates the selected dataset and prepares the
training/validation inputs inside the run workspace (
run/shared/). Your sourcedatasets/files are read-only and are never modified. - Iterative Development - The agent runs exploration, training, and evaluation cycles, scoring each iteration on the validation metric.
- Best-Model Snapshot - The best-performing iteration is tracked and copied
to
best_iteration_snapshot/. - Report Generation - Per-iteration and final reports are written to
reports/markdown/andreports/pdf/.
Results are written to the run's host workspace, outputs/<agent_id>/ by
default, which is mounted at /workspace in the container. See
Understanding Outputs for the full layout.
To score a finished run against a labeled held-out set, run inference with
--label-col; see Running Inference.
Monitoring Progress¶
During execution you'll see:
- Current iteration number
- Agent step (exploration, training, etc.)
- Validation metrics after each iteration
- Best-iteration tracking
Stopping a Run¶
Press Ctrl+C to stop. The agent attempts to save current progress before
exiting.
Next Steps¶
- CLI Options - Complete flag reference
- Understanding Outputs - What the agent produces
- Custom Prompts - Customize agent behavior