Retention Viability: High Viability (Daily Keeper)
Dominant Friction Factor
None - tight feedback loop with deterministic fallbacks
Execution Step Trace
[System] Workbench initialized. Preset loaded.
Agent Viability & Task Friction Dynamics
Read the explanation
The workbench calculates daily agent utility by penalizing latency friction, human intervention burden, and context decay risk. Restricting the context window to four thousand tokens with raw memory saturates capacity to ninety-two percent, triggering context decay penalties. Injecting a synthetic fault or failing tool calls drops utility below fifty, shifting the retention viability to high abandonment.
AI Agent Viability & Task Friction Workbench
How does this workbench calculate agent utility and retention viability from latency, human intervention thresholds, and context window settings?
This interactive workbench models autonomous AI agent workflows using a D3 force-directed graph alongside an educational scoring model for daily retention. It calculates an illustrative Daily Utility Score (capped between 10 and 100) by subtracting penalties for estimated latency friction, human-in-the-loop (HITL) intervention overhead, and context window saturation. When synthetic tool faults or unbuffered context limits occur, the model applies heavy friction penalties that downgrade retention viability from daily adoption to abandonment.
The Daily Utility Score, latency benchmarks, and context saturation levels are synthetic educational heuristics rather than empirically validated production benchmarks. Latencies and step counts are static fixtures assigned to presets, and context decay risk is a piecewise threshold lookup rather than a measurement of live LLM token attention degradation.
Try a worked example
From the default 'Contextual File Organizer' preset (utility score 88, latency 2500 ms, HITL intervention 12%, low decay risk), click 'Inject Synthetic Fault'. The Context Decay Risk immediately shifts to 'CRITICAL FAIL', the context meter fills to 100%, and the Daily Utility Score drops to 10, changing the retention badge to 'Low Viability (High Abandonment)'. Clicking 'Reset' restores the default topology, clearing the fault flag and returning the utility score to 88.
Friction and Utility Score Computation
The workbench evaluates daily agent retention via an illustrative formula: Daily Utility = 100 - min(35, (baseLatency / 1000) * 4) - (hitlRate * 0.6) - decayPenalty. Scores of 75 and above are classified as 'High Viability (Daily Keeper)', 50 to 74 as 'Moderate Viability (Occasional Tool)', and below 50 as 'Low Viability (High Abandonment)'.
Human-in-the-loop (HITL) intervention rate begins at (1.0 - threshold) * 100 percent, adding 15 percent when the error strategy is configured to prompt a human operator. Estimated base latency averages preset benchmark task timings, adding an 800 ms penalty when tool calls retry with exponential backoff.
Context Decay and Fault Injection Logic
Context window saturation is modeled as a categorical lookup: context windows of 4k tokens with raw memory trigger a 92% saturation and a 30-point decay penalty, while 32k tokens trigger moderate risk (65% saturation, 15-point penalty), and windows of 128k or higher remain at low risk (22% saturation, zero penalty).
Arming the 'Inject Synthetic Fault' button raises context saturation to 100%, triggers an extra 40-point decay penalty, and marks the dominant abandonment factor as an unrecoverable tool calling fault, illustrating how runtime fragility disrupts autonomous workflows.