AI Agent Workflow Simulator

Execution Workbench & Context Payload Inspector

Preset:
1x
DAG Node Topology Canvas
6 Nodes 5 Connections
Router LLM Tool Call Guardrail Human Gate
Click any node to inspect payload state

Interactive Parameter & Guardrail Tuning

$50

Triggers Human Gate when refund > threshold

Forces API lookup to trigger fallback self-correction

Manual decision state when queue is reached

Execution Telemetry IDLE

TOTAL STEPS 0 / 6
TOKENS CONSUMED 0
LATENCY (MS) 0 ms
ACTIVE NODE None

Context Payload Inspector

Node Payload
Input Prompt / Context:
I was charged twice for order #8492 and need a refund immediately.
Evaluated Memory State: 
{
  "order_id": "#8492",
  "intent": "refund_request",
  "amount_detected": 75.00
}
Step Output Decision:
Awaiting execution start...
Execution Log
[System] Workbench loaded. Selected preset: Customer Support Triage.
Verified Execution Record

Workflow Completion Artifact COMPLETED_WITH_HUMAN_GATE

EXECUTION ROUTE
Input Router → Intent Classifier → Account Lookup → Policy Evaluator → Human Escalation Gate → Customer Reply
OBSERVED RESULT & BOTTLENECK
Refund request verified ($75.00). Escalated to billing approval queue due to $50 threshold policy rule.
STATE EXPORT DATA HANDOFF
Ready for JSON state serialization.

What a fixed agent trace does and does not prove

Read the explanation

The default refund story advances through six stored entries. Their token counts sum to fourteen hundred twenty, and their latency counts sum to fifteen hundred seventy five milliseconds. These are literals in the preset, not observed model or account lookup measurements. The final entry says seventy five dollars was credited, but this page calls no payment service and transfers no funds. Step and Run reveal predefined outputs and accumulate their assigned telemetry. The policy entry starts with a seventy five dollar amount and a fifty dollar threshold. Moving the threshold slider to one hundred updates the rule threshold field. It does not recompute the requires human approval flag, change the stored output, or skip the human node. The supervisor decision is already the literal APPROVE. The human gate selector has no event handler in this source. These controls illustrate parameters without implementing actual approval enforcement. At one times speed, Run reveals one step every twelve hundred milliseconds. Six steps therefore take about seventy two hundred milliseconds to appear, while the cumulative latency display is fifteen hundred seventy five. Higher speed changes only the timer interval. A subsequent completion check ends the timer; none of this runs the described tools. Tool failure toggles one output string to a database timeout, while the array still proceeds toward the stored refund success. The research graph has two branches, but execution visits its step array sequentially and adds their latency, rather than launching real parallel workers. The reflection graph draws a retry edge, while its five entry array reaches final code without executing a second analysis pass. Export records the local preset and current state. Neither export proves a refund, compiled code, clinical research, or regulatory compliance. Offline checks preserve the same unavailable Cytoscape and style dependencies, not full working execution.

Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.