Workflow Execution

Execute a planned workflow node by node, tracking progress and handling failures.

Overview

You execute workflows by:

1. Reading the workflow state and plan
2. For each node, running the experiment directly using run_experiment
3. Analyzing results with vlm_inspect when visual verification is needed
4. Updating workflow state using workflow(action="update", ...)
5. Logging events using workflow(action="log", ...)
6. Making decisions on failures (or pausing for human input)

Safe State Updates

Use the workflow tool for all state changes. Never edit workflow.json directly.

Update Action

Updates workflow state with automatic validation:

workflow(
    action="update",
    workflow_id="calibrate_q0",
    data={
        "status": "running",
        "nodes.node_1.state": "success",
        "nodes.node_1.extracted": {"resonator_frequency": 5.823},
        "context.resonator_frequency": 5.823
    }
)

Features:

• Validates before writing (reverts on failure)
• Supports dot notation for nested updates
• Creates intermediate objects as needed
• Atomically updates multiple fields

Dot notation examples:

• "status": "running" - Simple field
• "context.resonator_freq": 5.823 - Nested path
• "nodes.node_1.state": "success" - Node by ID
• "nodes.node_1.experiment_id": "20240315_..." - Link node to experiment
• "nodes.node_1.extracted": {...} - Node nested field

Log Action

Appends events to history.jsonl with auto-timestamp:

workflow(
    action="log",
    workflow_id="calibrate_q0",
    event="node_completed",
    node="node_1",
    state="success",
    extracted={"resonator_frequency": 5.823}
)

The timestamp is added automatically.

Storage Structure

data/workflows/{workflow_id}/
├── workflow.json      # Current state (updated via workflow tool)
├── plan.md            # Planning document (read-only during execution)
└── history.jsonl      # Event log (appended via workflow tool)

Workflow State

workflow.json during execution:

{
  "id": "calibrate_q0",
  "name": "Calibrate Qubit Q0",
  "status": "running",
  "current_node": "node_2",
  "started_at": "2024-03-15T10:30:00Z",
  "nodes": [
    {
      "id": "node_1",
      "name": "Resonator Spectroscopy",
      "dependencies": [],
      "state": "success",
      "run_count": 1,
      "completed_at": "2024-03-15T10:32:00Z",
      "experiment_id": "20240315_103015_resonator_spectroscopy",
      "extracted": {"resonator_frequency": 5.823}
    },
    {
      "id": "node_2",
      "name": "Qubit Spectroscopy",
      "dependencies": ["node_1"],
      "state": "running",
      "run_count": 1,
      "started_at": "2024-03-15T10:32:15Z"
    },
    {
      "id": "node_3",
      "name": "Rabi Oscillations",
      "dependencies": ["node_2"],
      "state": "pending"
    }
  ],
  "context": {
    "resonator_frequency": 5.823
  }
}

Status values: created → running → paused / completed / failed

Node states: pending → running → success / failed / skipped

Execution Steps

Step 1: Load State

Read workflow.json using workflow(action="status", workflow_id="..."):

• Current workflow status
• Which nodes are complete (and their extracted values)
• Which node to execute next

Step 2: Find Next Node

Find a node where:

• state is pending
• All dependencies have state: success

If no such node exists:

• All nodes complete → mark workflow completed
• Some nodes failed/skipped with no path forward → mark workflow failed

Step 3: Execute Node

1. Mark node as running:

   workflow(
       action="update",
       workflow_id="calibrate_q0",
       data={
           "current_node": "node_1",
           "nodes.node_1.state": "running",
           "nodes.node_1.run_count": 1,
           "nodes.node_1.started_at": "2024-03-15T10:30:00Z"
       }
   )

2. Log to history:

   workflow(action="log", workflow_id="calibrate_q0", event="node_started", node="node_1", run=1)

3. Gather context from completed nodes and plan.md
4. Run the experiment using run_experiment:

   run_experiment(
     experiment_name="resonator_spectroscopy",
     params={
       "center_freq": 6.0,
       "span": 0.2,
       "num_points": 101,
       "power": -20,
       "num_averages": 2000
     }
   )

5. Analyze results using vlm_inspect when visual verification is needed:

   vlm_inspect(
     experiment_id="20240315_103015_resonator",
     prompt="Is there a clear dip? What is the resonator frequency?"
   )

6. Evaluate success criteria:
   • Check returned values (SNR, frequency, etc.)
   • Use VLM analysis for visual criteria
   • If criteria not met, retry with adjusted parameters (up to max attempts)
7. Proceed to Step 4 with the result

Step 4: Evaluate Result

On success:

1. Extract values as specified in plan.md
2. Update state (always include experiment_id from the run_experiment result's id field — this links the node to its plot in the UI):

   workflow(
       action="update",
       workflow_id="calibrate_q0",
       data={
           "nodes.node_1.state": "success",
           "nodes.node_1.completed_at": "2024-03-15T10:32:00Z",
           "nodes.node_1.experiment_id": "20240315_103015_resonator_spectroscopy",
           "nodes.node_1.extracted": {"resonator_frequency": 5.823},
           "context.resonator_frequency": 5.823
       }
   )

3. Log completion:

   workflow(
       action="log",
       workflow_id="calibrate_q0",
       event="node_completed",
       node="node_1",
       state="success",
       extracted={"resonator_frequency": 5.823}
   )

4. Go to Step 2 (next node)

On failure (after max attempts):

1. Update state:

   workflow(
       action="update",
       workflow_id="calibrate_q0",
       data={
           "nodes.node_2.state": "failed",
           "nodes.node_2.completed_at": "2024-03-15T10:33:00Z",
           "nodes.node_2.observations": "No peak found after 3 attempts"
       }
   )

2. Log failure:

   workflow(
       action="log",
       workflow_id="calibrate_q0",
       event="node_completed",
       node="node_2",
       state="failed",
       observations="No peak found after 3 attempts"
   )

3. Go to Step 5 (decide next action)

Step 5: Handle Failure

Review the failure and history. Decide:

If clear what to do (based on your analysis):

• Backtrack: A previous node's result may be invalid

  # Reset nodes
  workflow(
      action="update",
      workflow_id="calibrate_q0",
      data={
          "nodes.node_1.state": "pending",
          "nodes.node_2.state": "pending"
      }
  )
  # Log backtrack
  workflow(action="log", workflow_id="calibrate_q0", event="backtrack", from_node="node_2", to_node="node_1", reason="Resonator may have drifted")

• Skip: Node is optional, continue without it

  workflow(
      action="update",
      workflow_id="calibrate_q0",
      data={"nodes.node_2.state": "skipped"}
  )
  workflow(action="log", workflow_id="calibrate_q0", event="node_skipped", node="node_2", reason="Optional measurement")

If unclear, pause for human:

1. Update workflow status:

   workflow(
       action="update",
       workflow_id="calibrate_q0",
       data={
           "status": "paused",
           "paused_at": "2024-03-15T10:34:00Z",
           "paused_reason": "Qubit spectroscopy failed after 3 attempts",
           "suggestions": [
               {"action": "backtrack", "target": "node_1", "reason": "Resonator frequency may have drifted"},
               {"action": "retry", "reason": "Try with wider frequency range"},
               {"action": "skip", "reason": "Continue without qubit frequency"},
               {"action": "abort", "reason": "Stop workflow, investigate hardware"}
           ]
       }
   )

2. Log pause:

   workflow(action="log", workflow_id="calibrate_q0", event="workflow_paused", reason="Qubit spectroscopy failed")

3. Inform user and wait for decision

Resuming a Paused Workflow

When user provides a decision:

1. Log the decision:

   workflow(action="log", workflow_id="calibrate_q0", event="human_decision", decision="backtrack", target="node_1")

2. Execute the decision:
   • backtrack: Reset nodes and continue from target
   • retry: Re-run current node with new approach
   • skip: Mark node skipped, continue
   • abort: Mark workflow failed, stop
3. Update status:

   workflow(action="update", workflow_id="calibrate_q0", data={"status": "running"})

4. Continue execution from Step 2

Resuming After Interruption

If workflow shows status: running but was interrupted:

1. Check current_node - this node was interrupted
2. Treat it as a fresh start for that node (reset to pending)
3. Continue execution normally

Checking Status

Use the workflow tool:

workflow(action="status", workflow_id="calibrate_q0")

Shows:

• Current status and progress
• Node states
• Recent history
• Suggestions (if paused)

Example Execution Session

User: Execute workflow calibrate_q0

Agent: Starting workflow "Calibrate Qubit Q0"...

[Calls: workflow(action="status", workflow_id="calibrate_q0")]

Executing node_1: Resonator Spectroscopy...

[Calls: workflow(action="update", ..., data={"current_node": "node_1", "nodes.node_1.state": "running", ...})]
[Calls: workflow(action="log", ..., event="node_started", node="node_1")]

> Running experiment with center_freq=6.0 GHz, span=0.2 GHz
> Experiment complete: 20240315_103015_resonator
> Analyzing plot with vlm_inspect...
> VLM: "Clear dip visible at 5.823 GHz. SNR ~12."

Node_1 SUCCESS
- Extracted: resonator_frequency = 5.823 GHz

[Calls: workflow(action="update", ..., data={"nodes.node_1.state": "success", "nodes.node_1.extracted": {...}, "context.resonator_frequency": 5.823})]
[Calls: workflow(action="log", ..., event="node_completed", state="success")]

Moving to next node...

Executing node_2: Qubit Spectroscopy...

> Using resonator_frequency = 5.823 GHz from context
> Attempt 1: center=4.5 GHz, span=0.2 GHz
> Analyzing plot... No clear peak found
> Attempt 2: Expanding range to span=0.4 GHz
> Analyzing plot... Still no peak visible
> Attempt 3: Expanding range to span=0.6 GHz
> Analyzing plot... No peak found

Node_2 FAILED after 3 attempts
- Observations: No peak found with increasing range

[Calls: workflow(action="update", ..., data={"nodes.node_2.state": "failed", "status": "paused", ...})]
[Calls: workflow(action="log", ..., event="node_completed", state="failed")]
[Calls: workflow(action="log", ..., event="workflow_paused")]

Pausing workflow for human decision.

Suggestions:
1. Backtrack to node_1: Resonator frequency may have drifted
2. Retry with different center frequency
3. Skip this node (will affect downstream calibration)
4. Abort workflow

Please advise which action to take.

User: Try backtracking to node_1

Agent: Backtracking to node_1...

[Calls: workflow(action="log", ..., event="human_decision", decision="backtrack")]
[Calls: workflow(action="update", ..., data={"nodes.node_1.state": "pending", "nodes.node_2.state": "pending", "status": "running"})]

Executing node_1: Resonator Spectroscopy (run 2)...

> Running experiment...
> Found dip at 5.819 GHz (shifted from 5.823)

Node_1 SUCCESS
- Extracted: resonator_frequency = 5.819 GHz

[Updates state, continues to node_2...]

Best Practices

1. Execute experiments directly - Use run_experiment and vlm_inspect tools
2. Use workflow tool for state changes - Never edit workflow.json directly
3. Log key events - Record node start, experiment completion, node completion
4. Use context from previous nodes - Pass extracted values to subsequent experiments
5. Be conservative with failures - When unsure, pause and ask the user

Workflow Tool Actions Reference