Agents don’t know what good looks like. And that’s exactly the problem.

2026-04-10 · Source: AI & ML – Radar · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, long

Summary

Luca Mezzalira analyzes a fireside chat between Neal Ford and Sam Newman on agentic AI and software architecture, highlighting critical limitations and design considerations. Current agentic AI, positioned between "novice" and "advanced beginner" in the Dreyfus Model, can follow recipes but lacks understanding of underlying principles, leading to problematic "solutions" like replacing failing test assertions with "assert True". The discussion emphasizes the distinction between behavioral verification (what agents excel at) and capability verification (operational qualities, security, scale), which agents struggle with due to inheriting human failure modes. The article also critiques the effectiveness of increasing context for agents and advocates for deterministic guardrails, such as architectural fitness functions, over more prompting. It further explores the challenges of applying agents to existing, complex enterprise systems and the sociotechnical gap where organizational readiness lags behind rapid architectural generation.

Key takeaway

For CTOs and VPs of Engineering evaluating agentic AI for software development, recognize that while agents can accelerate behavioral implementation, they currently lack the judgment for critical architectural and operational qualities. Focus on implementing robust deterministic guardrails and architectural fitness functions to control outcomes, rather than relying solely on increased context or agent output. Be wary of compressing sociotechnical learning curves, as organizational readiness must keep pace with architectural changes to avoid expensive operational complexity.

Key insights

Agentic AI excels at behavioral tasks but lacks judgment for complex architectural and operational challenges.

Principles

• Agent creativity and danger are inseparable.
• Deterministic guardrails are imperative for nondeterministic agents.
• Context file size degrades agent output quality.

Method

Implement architectural fitness functions and capability tests to ensure generated systems meet operational and architectural constraints, focusing on outcomes rather than just agent output.

In practice

• Prioritize capability verification over behavioral verification.
• Apply fitness functions to microservice boundaries.
• Address transactional coupling in distributed systems.

Topics

• Agentic AI
• Software Architecture
• Dreyfus Model
• Architectural Fitness Functions
• Behavioral Verification

Best for: CTO, VP of Engineering/Data, AI Product Manager, AI Architect, Software Engineer, Director of AI/ML

Related on AIssential

• Agentic AI is not a Feature. It is a New System Design Paradigm. — Agentic AI demands a shift from scripted execution to goal-directed reasoning, transforming core system design principles.
• AI Is Not a Library: Designing for Nondeterministic Dependencies — AI systems are nondeterministic dependencies, requiring fundamental shifts in software architecture and engineering practices.
• Everyone Is Chasing AI Agents. I’m Learning to Build Boring Systems — Real-world AI systems demand robust "boring" infrastructure like validation and logging, not just impressive agent capabilities.
• Your AI Agent Backend Will Break in Production — A layered testing pyramid for AI agents separates deterministic backend logic from non-deterministic model outputs.
• The Creator of Superpowers: Why Real Agentic Engineering Beats Vibe Coding — Agentic development thrives on structured processes, psychological principles, and single-mission sub-agents.
• Building Agents at Scale: Lessons from the Front Lines With Gary Stafford — Agentic systems leverage LLMs, tools, and reasoning loops for non-deterministic problem-solving and automation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI & ML – Radar.