Cybernetics and the “human-on-the-loop” in agentic coding

2026-04-20 · Source: Thoughtworks Insights · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Advanced, long

Summary

The article "Cybernetics and the "human-on-the-loop" in agentic coding" by Dirk Lässig, published April 20, 2026, proposes a shift from "human-in-the-loop" to "human-on-the-loop" in software development due to the high velocity of AI-generated code. Traditional line-by-line review creates a bottleneck, necessitating a new steering role for humans. The author draws parallels between leading agentic SDLCs and managing organizations, advocating for the adoption of management practices. Cybernetics, introduced by Norbert Wiener in the 1940s and further developed by Stafford Beer with the Viable System Model (VSM) in the 1970s, provides a framework for understanding and controlling complex systems like AI agent teams. This approach involves humans operating at a "meta level" to design and configure agentic systems, using mechanisms like attenuation and amplification of variety (harness engineering) to achieve homeostasis. The Lean management practice of "Go See" (Gemba) is also integrated, requiring engineers to periodically perform manual spot-checks to update their system model and maintain technical intuition.

Key takeaway

For AI Architects designing agentic coding systems, you must transition from direct code review to a "human-on-the-loop" steering model. This means focusing on meta-level system design, defining objectives, and implementing cybernetic controls like variety attenuation and amplification. Regularly perform "Go See" spot-checks to validate agent behavior and update your system model, ensuring agents produce reliable, consistent results without becoming a bottleneck.

Key insights

Effective agentic coding requires humans to shift from line-by-line review to meta-level steering, applying cybernetic and management principles.

Principles

• LLM agents are non-deterministic.
• Requisite variety is essential for control.
• Effective regulators must model their system.

Method

Lead agentic SDLCs by designing the system, setting objectives, establishing boundaries, and intervening only for exceptions, rather than reviewing every code change.

In practice

• Configure dashboards for agent metrics.
• Encode decision rules in agent policies.
• Conduct weekly manual code spot-checks.

Topics

• Agentic Coding
• Cybernetics
• Human-on-the-Loop
• SDLC Management
• Viable System Model
• Harness Engineering
• Lean Management

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

Related on AIssential

• Hi, I’m Venelin and I Use Claude Code | A Confession From Recovering User — Agentic coding tools automate the full software development cycle, but human oversight remains critical for validation.
• The MCP Architecture: Engineering the Automotive Cognitive Loop for High-Performance AI — MCP standardizes LLM interaction with local tools and data, enabling robust, autonomous AI agents.
• My Brief History of AI Coding — Agentic AI transforms coding from a tool-assisted task into a collaborative, exploratory process.
• Evaluating AI agents: Real-world lessons from building agentic systems at Amazon — Evaluating agentic AI requires a holistic framework assessing emergent system behaviors beyond individual LLM performance.
• Agents: Inner Loop vs Outer Loop — Agent "loop closing" is about learned behavior within a hardcoded cycle, not new infrastructure.
• Autoresearch, Agent Loops and the Future of Work — AI agentic loops enable autonomous, iterative experimentation and continuous improvement across various domains.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Thoughtworks Insights.