Agentic system harnesses structure multi-model workloads as directed graphs, breaking complex prompts into coordinated parallel execution tasks. Outputs route into peer evaluation nodes. Verifier strictness sets the statistical pass threshold, validating claims before allowing downstream consensus. When verification fails, the pipeline routes into a self-correction loop, spending additional simulated tokens and latency to refine results up to the retry limit. Clicking Run Loop triggers consecutive execution steps, updating the waterfall timeline, token cost telemetry, and final verified pass rate.
How does this simulator model multi-agent pipelines, token costs, and verification checks?
This simulator visualizes agent topologies as directed graphs across three predefined configurations: research, code sandbox, and consensus. Rather than issuing live API calls or parsing actual agent outputs, each step draws random token counts (250–650 tokens) and generates synthetic execution latencies (0.2–0.6 seconds per node). Nodes advance sequentially in array index order, accumulating approximate costs based on fixed rates ($0.003 or $0.0005 per 1,000 tokens) while evaluating pass/fail status using a pseudo-random check against the strictness threshold.
The simulator is a pedagogical front-end demonstration: it does not connect to real language model APIs, perform code execution, or evaluate factual accuracy. Graph traversal processes nodes sequentially by list order rather than resolving true asynchronous DAG edge dependencies or branching dynamically based on evaluator pass/fail results. Token prices and latencies are hardcoded synthetic random variables rather than empirical benchmark measurements.
Select the 'Code Synthesis & Sandbox Test' option from the Pipeline Topology dropdown. The active node resets, clearing prior execution logs and loading 5 pipeline nodes (Spec Architect, Module Coder, Test Sandbox, AST Self-Corrector, and Deployer). Clicking 'Step' advances execution to the first node (architect), logging an execution role event and incrementing the total simulated tokens, cumulative latency, and dollar cost.
Although the visualizer uses Cytoscape to render nodes and directed edges (such as conditional fail branches to recovery nodes), the runtime stepping function iterates through nodes using an array index counter (stepIndex). Each press of the Step button evaluates the next node in the preset list sequentially, resetting previous executing states to complete.
Token usage is generated per step as Math.floor(250 + Math.random() * 400). Token cost scales by $0.003 per 1,000 tokens if the node's assigned model is claude-3-5-sonnet, or $0.0005 per 1,000 tokens otherwise. When an evaluator node executes, a pseudo-random draw against the verifier strictness slider determines whether a PASS or FAIL log entry is recorded, updating the pass rate percentage.