Claims about AI productivity improvements

I have personally found huge personal gains on my personal project (where I have complete control) and low-medium significant gains at my job in terms of productivity by REJECTING the 'agentic' workflow premise.

The main problems I see on people not having success with AI are the following:

- Not spending enough time on understanding how to prompt properly, and configure your setup to contextualize the AI properly i.e. Markdown files that: summarize your project structure, explain backend or frontend workflows, business logic, and design decisions, coding standards, (CLAUDE.md for CC users) where you can easily tell the AI to read and they will code how you want.

- Check every single LLM output and patch suggestion with the SAME cognitive load you would use to actually coding it yourself. This is the most important, or else you are comparing apples to oranges.

- Context Engineering: Using subagents to find out how a function or pipeline works end to end to feed your main agent with a succinct summary, keeping the main coding agent on track as multiple diverting tasks poisons the context and effectiveness massively.

- Ask for a sub Agent to verify the work given the spec, with a goal for maintainability, scalability and security.

- Linting (with strict standards), formatting and testing rigorously ( I have them as pre-commits and forbid any commits that have a single linting issue or less than 80% test coverage (if applicable).

Following this I have had massive successes for the simple reason the LLM can write code way faster than I could possibly type. For me this is the main productivity gain of LLMs if you have it set up properly, it can be a massive autocomplete, where if correctly enforced and contextualized it can make huge productivity gains because it can simply write code multitude times faster than I could possibly physically type, inherently making me more productive. This is someone with 90+ WPM using Vim.

Fully agentic autonomous workflows for me are a pipe dream and not feasible at all given due to silly optimizations that backfire, most notably wanting to preserve patch context windows when patching a file, and importing modules (for python) in the middle of the script, or making extremely silly workarounds for a simple syntax error.

If people took the time to set up the proper guardrails, gave it the same cognitive load as normal programming, hopefully they could see the massive boost I have seen, it truly is remarkable especially the more you understand and know the whole codebase because you can easily contextualize what it needs and it produces a solid first draft you just have to edit.

For these reasons, I take it with a massive grain of salt this article.