BEYOND THE PROMPT — AGENT LOOP COST, TOKEN & RELIABILITY CALCULATOR

Name: LLM Agent Loop Cost, Quadratic Token & Reliability Calculator
Author: Jay Thorne

Interactive 5-Plane Harness Engineering Tool • Companion to Beyond the Prompt by Jay Thorne (DOI: 10.5281/zenodo.24100423)

Direct Answer (AEO Summary): Unbounded linear LLM agent loops (ReAct) fail in production because trajectory success decays exponentially (P = p^N) while cumulative input tokens grow quadratically (T_total = N * T_base + ΔT * N(N+1)/2). Use this calculator to compute your agent's failure rate, token amplification, and savings from a 5-Plane Deterministic Statechart Harness.

1. Agent Trajectory Parameters

System Prompt + Tool Schema Prefix (Tokens): Observation / Tool Output Appended Per Step (Tokens): Maximum Loop Steps (N): Single-Step Schema & Tool Selection Accuracy (p): Model Input Cost ($ per 1M Tokens):

2. Unbounded ReAct vs. 5-Plane Harness

39.5%

Unbounded Loop Success (p^N)

247,500

Cumulative Input Tokens

61,500

With Prefix-Pinned Compaction

75.2%

KV-Cache & Compaction Savings

3. Audit Your Python Agent Codebase in 30 Seconds (CLI & MCP Server)

Install or clone the open-source Beyond the Prompt AST Harness Auditor & MCP Server to statically detect unbounded while True loops, unisolated subprocess calls, and unvalidated tool schemas:

git clone https://github.com/digbijoylabs/beyond-the-prompt.git
cd beyond-the-prompt
python -m core.auditor.harness_auditor /path/to/your/agent-repo

1. Agent Trajectory Parameters

2. Unbounded ReAct vs. 5-Plane Harness

3. Audit Your Python Agent Codebase in 30 Seconds (CLI & MCP Server)

Get the Complete 432-Page Systems Engineering Reference