How AI Code Assistants Are Redefining the Developer’s Role in 2026

Lee Robinson and Boris Cherny recently shared their AI insights, with Lee’s talk “AI codes better than me. Now what?” illustrating how developers need to reposition themselves starting in 2026.

Robinson notes that he now rarely writes code himself; for example, he used AI to implement a Rust image‑compression tool compiled to WebAssembly and run in a SvelteKit app without touching a single line of code. He also built a Space Invaders clone for the Analog Pocket handheld using Verilog, a language he knew nothing about, thanks to AI assistance.

These examples show AI’s ability to break traditional technical silos, prompting Robinson to join Cursor to help shape the transformation of software engineering.

“AI will no longer be confined to a single technology tag.”

Current models such as Codex Max and Opus 4.5 already outperform human developers in code generation, leading Robinson to advise that developers should stop comparing their coding ability with AI.

“Therefore, developers should not get hung up on ‘writing code’ because you simply can’t beat AI at that.”

Robinson’s recommendations for developers are:

Don’t assume AI can’t improve in a domain just because it currently struggles; model progress is rapid.

Let AI take over repetitive and tedious development tasks.

When generated code becomes usable, engineer value shifts to “taste” – focusing on product, user experience, marketing, and building things users love.

“The future core is no longer ‘how to write code’ but ‘how to build and manage great products’ using AI.”

Even tools like Android Studio’s new AI Agent are adding vibe‑coding support, making AI an indispensable part of the developer ecosystem, similar to how StackOverflow’s traffic has plummeted as AI takes over.

Manus, recently acquired by Meta, targets non‑technical users who care only about results, suggesting that AI‑generated prototypes may reduce the importance of the coding process itself.

“If AI can produce an initial result to validate a product, the intermediate coding steps may become less critical.”

A Google Gemini API lead demonstrated that Claude Code could reproduce a year’s worth of team work in one hour, underscoring that comparing code output with AI is futile.

“Don’t compare AI’s code output ability; instead, focus on how you use it.”

Boris Cherny reported that in 30 days he ran 259 pull requests entirely with Claude Code + Opus 4.5, with the longest session lasting nearly 42 hours. He shared practical tips for getting the most out of Claude Code:

Run five Claude processes in parallel, label them 1‑5, and use system notifications to identify which process needs human attention.

Maintain a shared CLAUDE.md file where the team records mistakes for Claude to avoid in future runs.

During code review, use @.claude to automatically update rules in CLAUDE.md via a Claude Code GitHub Action.

Start most sessions in “plan” mode (Shift+Tab twice); a good plan prevents blind, unfocused work.

Package frequent workflows as slash commands stored in .claude/commands/ so Claude can invoke them autonomously.

Avoid the --dangerously-skip-permissions flag; instead, pre‑approve safe bash commands with the /permissions option and store the configuration in .claude/settings.json.

Use the ralph-wiggum plugin for long‑running tasks, allowing Claude to loop until completion.

Provide Claude with a way to verify execution results; feedback mechanisms (e.g., Claude Chrome extension) can improve output quality by 2‑3×.

These systematic configurations and workflow rules make AI more intelligent, intuitive, and maintainable, turning AI management into an essential skill for future developers – perhaps even a new role of “AI manager.”