The Outside of the Loop

2026-06-14 · Source: Machine Learning on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, long

Summary

AI coding assistants like Claude Code are rapidly advancing, with the author reporting 80% of their coding skills are now auto-written, dramatically improving workflows. Anthropic internally states that 80% of its production code is written by Claude, enabling engineers to ship eight times more code per quarter. Models that once handled four-minute tasks now run for twelve hours unattended, and their ability for "discernment"—picking optimal next steps—is surpassing human judgment, reaching 64% by April 2026 in research step selection. Despite this progress, the article asserts that two crucial human roles remain "outside the loop": "choosing the problem" (defining the initial objective) and "judging the result" (subjectively evaluating the outcome). It warns that while objective verification is being automated, subjective "taste" remains human, leading to a "barbell" organizational structure where the middle layer of producers is removed, posing a challenge for developing future talent.

Key takeaway

For Directors of AI/ML navigating rapid automation, recognize that while AI excels at production and objective discernment, your critical role shifts to defining strategic problems and subjectively judging outcomes. You must actively cultivate "taste" in your teams by protecting opportunities for junior talent to engage in "slow work" and learn judgment, even if it appears inefficient. Failing to do so risks a future talent gap in critical decision-making and quality assessment.

Key insights

AI excels at execution and discernment within a frame, but humans must define problems and subjectively judge outcomes.

Principles

• AI systems self-improve by iterating on variations.
• Human "taste" defines problems and judges subjective quality.
• Objective verification is automatable; subjective judgment is not.

Method

The article describes a self-learning loop where Claude Code analyzes user sessions, drafts new skills, and wires them, with human review for quality and alignment.

In practice

• Design loops that prompt coding agents, not direct prompts.
• Protect "slow work" for training human judgment.
• Focus on problem definition and subjective outcome evaluation.

Topics

• AI Agents
• Recursive Self-Improvement
• Human-AI Teaming
• Organizational Transformation
• AI Ethics
• Judgment and Taste

Code references

• NousResearch/hermes-agent

Best for: CTO, VP of Engineering/Data, Executive, AI Engineer, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• The AI Fog & The New Judgment Layer — AGaaS and Agent Expansion are shifting human work to a new "judgment layer" focused on oversight and critical decisions.
• Humans vs Machines: Who Wins Tomorrow? — AI excels at predictable, repetitive tasks but struggles with human-centric complexities like emotion, intuition, and creativity.
• Overconfidence at Scale — AI scales decisions exponentially, but human responsibility does not, creating systemic risks from scaled, flawed assumptions.
• Your AI Workflow Is Quietly Replacing Your Judgment. Here’s the Fix — The default AI workflow subtly erodes human judgment by training users to skip critical review.
• Anthropic Sues the Pentagon — Adapting to AI's paradigm shift requires "unlearning" outdated mental models and continuously updating one's understanding.
• The professional judgment gap: Tracing AI’s impact from lecture hall to professional services — AI adoption risks creating a "professional judgment gap" by automating entry-level experiences vital for skill development.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning on Medium.