Agentic
Human-in-the-loop ML pipeline management.
All methods are accessed via client.agentic.
start_run() Start a new agentic ML training pipeline run. This initiates the full ML workflow: data preparation -> label selection -> feature engineering -> model training -> deployment -> reporting -> monitoring. The pipeline will pause at each phase listed in require_approval and wait for a decision via submit_decision().
Parameters
model_name str Required
Name for the model being trained
model_description str default: ''
Description of the model's purpose
auto_mode bool default: False
If True, automatically proceed through phases without approval
require_approval list default: None
List of phases requiring human approval. Defaults to all phases. Valid phases: label_selection, data_preparation, feature_engineering, model_training, model_deployment, report_creation, monitoring_creation
auto_apply_safe_features bool default: True
Automatically apply safe feature engineering suggestions
auto_deploy bool default: False
Automatically deploy the model after training
run_id str default: None
Optional existing run ID to resume
user_query str default: None
Optional natural language description of what to build
phase_plan list default: None
Optional ordered list of phases to execute
Returns
StartRunResponse — StartRunResponse with run_id and status
Example
1 result = client . agentic . start_run (
2 model_name = "My Model" ,
3 model_description = "Predicts customer churn" ,
4 auto_mode = False ,
5 require_approval = [ ] ,
6 auto_apply_safe_features = True ,
7 auto_deploy = False
8 )
get_run_state() GET /v1/agentic/runs/ {run_id} Get the current state of an agentic pipeline run. Use this to poll for run status and see current phase, progress, and results.
Parameters
run_id str Required
The run ID returned from start_run
Returns
dict — Dict with run state including current phase, status, and any results
Example
1 result = client . agentic . get_run_state (
2 run_id = "run_abc123"
3 )
get_pending_decision() GET /v1/agentic/runs/ {run_id} /pending-decision Check if the pipeline is waiting for a human decision. When the pipeline reaches a phase listed in require_approval, it pauses and creates a pending decision. Call this to see what decision is needed, then use submit_decision() to provide your choice.
Parameters
run_id str Required
The run ID
Returns
dict — Decision info dict with decision_type, options, and context, or None if no decision pending
Example
1 result = client . agentic . get_pending_decision (
2 run_id = "run_abc123"
3 )
submit_decision() POST /v1/agentic/runs/ {run_id} /decisions Submit a decision for a pending approval in the pipeline. The required fields depend on the decision_type from get_pending_decision(). Common patterns: - Label selection: decision_type="label_selection", choice_index=N - Feature engineering: decision_type="feature_engineering", apply_indices=[...], skip_indices=[...] - Model training: decision_type="model_training", choice_index=N - Deployment: decision_type="model_deployment", action="approve" or "skip" - Report: decision_type="report_creation", report_config={...} - Monitoring: decision_type="monitoring_creation", monitoring_config={...}
Parameters
run_id str Required
The run ID
decision_type str Required
Type of decision (matches pending decision's decision_type)
choice_index int default: None
Index of chosen option from the options list
action str default: None
Action string (e.g., "approve", "skip", "reject")
custom_value any default: None
Custom value for decisions that accept free-form input
label_type str default: None
Label type for label selection decisions
apply_indices list default: None
Indices of items to apply (e.g., feature engineering steps)
skip_indices list default: None
Indices of items to skip
selected_indices list default: None
Indices of selected items
selected_options list default: None
List of selected option values
selected_features list default: None
List of selected features
skipped bool default: None
Whether to skip this phase entirely
done bool default: None
Whether the decision process is complete
report_config dict default: None
Configuration for report creation
monitoring_config dict default: None
Configuration for monitoring setup
Returns
SubmitDecisionResponse — SubmitDecisionResponse with status
Example
1 result = client . agentic . submit_decision (
2 run_id = "run_abc123" ,
3 decision_type = "label_selection" ,
4 choice_index = 1 ,
5 action = "..." ,
6 custom_value = . . . ,
7 label_type = "..."
8 )
get_phases() GET /v1/agentic/runs/ {run_id} /phases Get the phase execution history for a run.
Parameters
run_id str Required
The run ID
Returns
list — List of phase status dicts with phase name, status, timestamps, and any errors
Example
1 result = client . agentic . get_phases (
2 run_id = "run_abc123"
3 )
cancel_run() POST /v1/agentic/runs/ {run_id} /cancel Cancel a running agentic pipeline.
Parameters
run_id str Required
The run ID to cancel
Returns
CancelRunResponse — CancelRunResponse with success status
Example
1 result = client . agentic . cancel_run (
2 run_id = "run_abc123"
3 )
skip_phase() POST /v1/agentic/runs/ {run_id} /skip-phase Skip the current phase and move to the next.
Parameters
run_id str Required
The run ID
Returns
SkipPhaseResponse — SkipPhaseResponse with skipped_phase and next_phase
Example
1 result = client . agentic . skip_phase (
2 run_id = "run_abc123"
3 )
resume_run() POST /v1/agentic/runs/ {run_id} /resume Resume a paused run (e.g., after a timed-out decision was auto-resolved).
Parameters
run_id str Required
The run ID to resume
Returns
dict — Resume result dict
Example
1 result = client . agentic . resume_run (
2 run_id = "run_abc123"
3 )
send_chat() POST /v1/agentic/runs/ {run_id} /chat Send a chat message to the agentic pipeline and get a response. Use this to ask questions about the current state of the run, request explanations, or provide additional context.
Parameters
run_id str Required
The run ID
content str Required
The message content
Returns
ChatMessageResponse — ChatMessageResponse with the pipeline's reply
Example
1 result = client . agentic . send_chat (
2 run_id = "run_abc123" ,
3 content = "How is the model performing?"
4 )
get_chat_history() GET /v1/agentic/runs/ {run_id} /chat Get the chat history for a pipeline run.
Parameters
run_id str Required
The run ID
Returns
list — List of chat messages with role, content, and timestamp
Example
1 result = client . agentic . get_chat_history (
2 run_id = "run_abc123"
3 )
retrain() POST /v1/agentic/runs/ {run_id} /retrain Retrain a completed run with optional new preprocessing steps or parameters.
Parameters
run_id str Required
The run ID of the completed run to retrain
params dict default: None
Optional dict of retraining parameters
Returns
dict — Retrain result dict
Example
1 result = client . agentic . retrain (
2 run_id = "run_abc123" ,
3 params = { }
4 )
get_preprocessing_dag() GET /v1/agentic/runs/ {run_id} /preprocessing/dag Get the preprocessing DAG (directed acyclic graph) visualisation data for a run. Shows how preprocessing steps are connected and their order of execution.
Parameters
run_id str Required
The run ID
Returns
dict — DAG data dict with nodes and edges
Example
1 result = client . agentic . get_preprocessing_dag (
2 run_id = "run_abc123"
3 )
get_column_lineage() GET /v1/agentic/runs/ {run_id} /preprocessing/lineage Get column lineage graph for the run's preprocessing. Shows how columns are transformed through the preprocessing pipeline.
Parameters
run_id str Required
The run ID
Returns
dict — Column lineage data dict
Example
1 result = client . agentic . get_column_lineage (
2 run_id = "run_abc123"
3 )
