AI Agent Retention & Workflow Reliability Lab

Agent Archetypes:

Empirical Retention Index

92.4 / 100

High Retention Habit

Reliability Boundary Safety

98.2%

Human Oversight Burden

Low (1 Gate)

Hallucination Safety Margin

98.0%

Execution Loop Directed Graph (D3 Canvas)

Click node to edit configuration

Node Configuration: Human Approval (High Impact Rules) GATE

Node Label

Max Retries

LLM Temperature

Gate Approval Threshold

Noise & Stress Parameters

API Transient Error Rate 5.0%

Context Drift Rate 1.0%

Hallucination Risk Rate 2.0%

Execution Trace Console

Diagnostic Analysis: Identified bottleneck in Human Approval Gate (Average Wait Time: 4.2m). High overall retention due to bounded tool calls and deterministic schema enforcement.

Node-by-Node Execution Diagnostics

Node ID	Step Label	Type	Avg Latency	Error / Retry Rate	Retention Impact

A local score combines assigned workflow assumptions

Your browser does not support video.

Read the explanation

The default code reviewer fixture contains five nodes and four directed edges, not an actual connected GitHub workflow. API error rate point zero five subtracts ten points and drift point zero one subtracts one point five from reliability, giving eighty eight point five. Hallucination point zero two subtracts five from one hundred, giving ninety five. At three pixels per score point the reliability bar measures two hundred sixty five point five, hallucination margin two hundred eighty five and oversight complement seventy five two hundred twenty five. These invented indices are not measured user retention, model accuracy or production error rates. The generic score weighs reliability by point forty five, hallucination margin by point thirty five, and oversight complement by point twenty. Default contributions are thirty nine point eight two five, thirty three point two five and fifteen. Two assigned retries across five nodes add point eight, summing eighty eight point eight seven five, displayed eighty eight point nine. The loaded fixture instead carries ninety two point four before a recalculation. At three pixels per score point fixture measures two hundred seventy seven point two, generic two hundred sixty six point six two five and their gap three point five two five measures ten point five seven five. Fixture text is not a measured retention benchmark. Raising assigned API errors from five to fifteen percent subtracts twenty extra reliability points. The score coefficient point forty five turns that into nine fewer points, changing generic eighty eight point eight seven five to seventy nine point eight seven five, displayed seventy nine point nine. At three pixels per point the baseline measures two hundred sixty six point six two five, changed two hundred thirty nine point six two five and reduction twenty seven. Native inline noise controls and actual JSON download work even though missing D three aborts graph initialization; the graph draw function safely returns without its initialized group, so native presets can still update state. The human wait four point two minutes is canned, not measured. No real PR comments, email, agent execution or empirical retention is proved.