AI workflow 2026 – 2 years later, what works, what doesn’t, Claude 200K in practice

Two years after I first wrote about AI coding tools, the landscape has changed completely. Context windows went from 8K to 200K+ tokens. Models that were unreliable on Magento code are now genuinely useful. Cursor IDE turned AI from a suggestion machine into a pair programmer. I revisit my 2024 assessment with two years of production experience and concrete workflow changes.

What changed since 2024

Claude with 200K context – how I use it

# What 200K tokens enables for Magento work:

# Paste an entire Magento module (typical: 15-50K tokens)
find app/code/Vendor/Module -name "*.php" -o -name "*.xml" | \
    xargs wc -l | tail -1
# 3,847 lines = ~15K tokens - fits easily in context

# Workflow 1: Architecture review
# "Review this module for Magento 2 antipatterns, missing null checks,
#  places where ObjectManager is used directly, and Service Contract violations"

# Workflow 2: Explain unknown code
# "Explain what this observer does and how it interacts with the checkout"

# Workflow 3: Write tests for a complete module
# "Write PHPUnit tests for all service classes in this module.
#  Use mocks for all Magento interfaces."

# Workflow 4: Migration planning
# "This module was written for Magento 2.3. What needs to change for 2.4.8?"

# What still fails:
# - Generating correct di.xml for complex virtual type hierarchies
# - Layout XML for unusual block positions
# - Specific Magento 2 version quirks (2.4.6 vs 2.4.8 differences)

Cursor IDE – the real workflow change

# Cursor features that changed my daily workflow:

# 1. Codebase chat: @codebase "How does the checkout totals collection work?"
# Cursor reads your entire project and answers from actual code

# 2. Multi-file edit: describe a change, Cursor edits all affected files
# "Add a new extension attribute 'supplier_code' to Order"
# -> Creates extension_attributes.xml, plugin, interface, setup patch, all at once

# 3. Apply diffs from AI: "Refactor this to use the Repository pattern"
# -> Shows a diff, you apply or reject changes file by file

# 4. Rules files: .cursorrules - project-specific AI instructions

# .cursorrules - Magento 2 project rules for Cursor
You are a Magento 2 developer assistant. Follow these rules:

1. Always use constructor injection (DI), never ObjectManager directly
2. Return Data Interfaces (e.g. ProductInterface), not concrete Models
3. Use Service Contracts for cross-module data access
4. Always declare strict_types=1 in PHP files
5. Prefer plugins over preferences when possible
6. Always set a pageSize when using SearchCriteriaBuilder
7. Use readonly classes for DTOs and Value Objects (PHP 8.2+)
8. Use property hooks for validated setters (PHP 8.4+)
9. Exceptions: use Magento\Framework\Exception\* types

Magento version: 2.4.8, PHP version: 8.5

Local models in 2026 – privacy-safe workflow

# qwen2.5-coder:32b is the recommended local model for PHP in 2026
# Quality: ~85% of Claude 3.5 Sonnet on PHP tasks
# Speed: ~2-3 tokens/second on MacBook M3 Pro (32GB RAM)
# Privacy: runs entirely locally, no data leaves the machine

ollama pull qwen2.5-coder:32b

# Use in Cursor: Settings -> Models -> Add Custom Model
# URL: http://localhost:11434/v1
# Model: qwen2.5-coder:32b

# Use cases for local model:
# - Client code covered by NDA
# - Security-sensitive code (auth, payment)
# - Proprietary business logic
# - Working offline or on poor connection

Honest assessment: what AI still does badly

<?php

// STILL UNRELIABLE in 2026:

// 1. Magento 2 DI configuration (di.xml)
// AI generates plausible-looking XML but often:
// - Wrong xsi:type for the data being injected
// - Missing MKSTREAM flag on queue configs
// - Incorrect virtualType inheritance

// 2. Complex plugin interactions
// "Write a plugin that modifies X but only when Y is true and Z is not set"
// AI generates something that looks right but fails in edge cases

// 3. Specific version incompatibilities
// "Is this code compatible with Magento 2.4.6 AND 2.4.8?"
// AI often misses deprecation differences between minor versions

// 4. Performance implications
// AI generates functionally correct code that has N+1 problems,
// missing cache warm-up, or fetches the same data multiple times.
// It fixes issues when told about them but doesn't proactively identify them.

// 5. Security vulnerabilities
// Missing input validation, SQL injection via raw queries,
// missing ACL checks on REST endpoints.
// ALWAYS review AI-generated code for security manually.

ROI summary – actual time savings

Summary

Two years of daily AI-assisted development produced one clear conclusion: AI is a genuine 2-3x productivity multiplier for a developer who already knows what they are doing. It does not replace expertise; it amplifies it. The 200K context window is the single biggest improvement – being able to show Claude an entire module and ask “what’s wrong with this” is categorically different from pasting individual files. The remaining failure modes are consistent: trust AI for boilerplate, scaffolding, and test generation; verify manually for security, DI configuration, and performance-sensitive code.