LAI #128: Developers Using AI Were 19% Slower. They Thought They Were 20% Faster.

2026-01-08 · Source: Learn AI Together · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Advanced, medium

Summary

LAI #128 highlights a significant perception gap in AI-assisted development. Developers using AI were 19% slower on complex tasks, despite believing they were 20% faster. This contrasts with Claude Code's creator using it for 100% of his code since November 2025, and his team shipping 95% agent-written code. The brief also covers DeepSeek V4's mHC for stabilizing 61-layer training. A Snowflake Cortex optimization yielded a 9-point accuracy gain and 26% cost reduction. Further topics include building stateful research agents on TensorLake to preserve VM state. Transformers fail at exact arithmetic due to Maclaurin series limitations. A multi-agent research engine architecture features centralized state and human checkpoints. A community poll indicates Claude Code is preferred by 61% of users over Codex.

Key takeaway

For AI Engineers evaluating developer productivity tools, recognize that your perceived speed gains with AI might mask actual slowdowns on complex tasks. Implement rigorous, objective performance metrics beyond subjective feeling. If you are building agentic AI systems, prioritize robust state tracking and recovery mechanisms. This prevents duplicated work and ensures reliable operation, especially with retries. Consider exploring multi-agent architectures with centralized state for enhanced control and efficiency.

Key insights

The perception of AI's impact on developer productivity often misaligns with actual performance, especially on complex tasks.

Principles

• Agent recovery from failure requires explicit state tracking.
• Transformers have inherent mathematical limits for exact arithmetic.
• Centralized state improves multi-agent system traceability.

Method

To recover from agent failures, track agent intent, tool interactions, side effects, and allowed next steps after each action, enabling safe resumption.

In practice

• Implement state tracking for agent retries to prevent duplicated work.
• Use Genetic-Pareto optimization for AI function tuning.
• Employ named sandboxes for stateful research agents.

Topics

• AI Developer Productivity
• Agentic AI Systems
• DeepSeek V4
• Transformer Architecture
• Snowflake Cortex AI
• Multi-Agent Systems
• State Management

Code references

• webxos/lack

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

Related on AIssential

• How I deleted 95% of my agent skills and got better results — Nick Nisi, WorkOS — Agent performance improves by enforcing actions with code, guiding with specific "gotchas," and rigorously measuring outcomes, not by comprehensive instruction.
• Ideas: slow down to speed up when working with AI agents — AI coding tools are rapidly adopted, prompting a "slow down to speed up" approach.
• 😺 Anthropic: AI Is Building AI now — AI systems are increasingly handling execution tasks in their own development, fundamentally shifting human roles.
• Breaking the Spell of Vibe Coding — Vibe coding can induce a "dark flow" state, similar to gambling, leading to misperceptions of productivity and skill development.
• Faster Code, Slower Understanding: A Survival Note from a Year of Agent-First Engineering — AI shifts the engineering bottleneck from code production to comprehension and maintenance, demanding new optimization strategies.
• Agentic AI Tackles RTL Verification’s Productivity Gap — Agentic AI enhances RTL verification productivity by coordinating complex workflows while keeping humans in control.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Learn AI Together.