AI Agents in Enterprise: Why Compound Errors and Token Costs Demand Architectural Rigor

Introduction: Calm Thinking Amid the Agent Hype

In 2025, AI Agents have become the buzzword for "AI-native applications," envisioned as autonomous digital employees that can analyze tasks and dynamically invoke tools. However, the hype demands a sober assessment: the real challenge for B2B agent systems is not model capability but whether the architectural approach can withstand the demands of complex enterprise systems.

1. The Mathematics of Compound Failure in Multi-Step Execution

Many assume that a sufficiently powerful model can autonomously plan and execute a multi-step workflow. The hidden engineering problem is uncontrollable compound error . Drawing on probability analysis from Utkarsh Kanwat's "Stop Believing the Agent Hype," the article illustrates:

If each step has a 95% reliability, the overall success rate after 10 steps drops to 59%.

Even at an ideal 99% per step, 20 steps yield only an 82% success rate.

Enterprise systems require 99.9%+ end-to-end stability.

The conclusion: the issue is not model intelligence but the reality that layered uncertainty cannot be tolerated in production engineering. Engineering fundamentals require containing every unreliable point within a structured system, not letting the model freely orchestrate workflows.

2. Context-Based Memory: A Technical and Economic Dead End

Another underestimated problem is the irrational token cost structure and state management of current agent designs. Many systems maintain "memory" by concatenating the full context every turn. This approach has two consequences:

Token costs grow near-exponentially with conversation turns, making large-scale deployment prohibitively expensive.

System state remains opaque and unabstracted—all variables are piled into the context, causing model confusion and poor recoverability.

This "let the model remember everything" mindset is essentially an absence of architectural responsibility . For agents to scale, the architecture must redesign how task state is organized and passed: stateless execution, functional expression, and explicit state modeling are necessary architectural shifts.

3. Architecture Contains AI, Not the Reverse

Echoing the author's earlier piece "The Truth of Agent Architecture: 90% Software Engineering, 10% AI," the core thesis is: agent systems are not AI-led architectures but architectures that accommodate AI.

An agent is not a replacement for the system; it is an intelligence interface for structure.

The LLM's role should be limited to:

Translating vague human intent into clear function-call intent.

Converting unstructured data (natural language, images) into structured parameters.

Decomposing complex goals into controllable, orchestratable units within the system.

Actual execution remains the job of stable, traceable, rollback-capable execution engines. Therefore, a usable agent system must achieve three architectural imperatives:

Explicit responsibility boundaries: AI only parses; it does not schedule.

Reliable execution: Every task must be definable, observable, and recoverable.

State decoupling: Task state flows through explicit structures, not context stacking.

Reference: Utkarsh Kanwat, "Stop Believing the Agent Hype."