Agentic Work in Real Life

LLMs and coding agents are changing software development. The harder question is what actually works, why it works, and where the limits are.

This series belongs to neither extreme of the AI debate. It is not AI advertising, an anti-AI argument, a model leaderboard, or a collection of supposedly perfect prompts. It examines real work with LLMs and coding agents, the mechanisms behind their behavior, and the consequences for professional software engineering.

A record from late 2026

Capabilities have changed rapidly since around 2023. A limit observed today may become outdated quickly, so concrete models, prices, and benchmark results appear only as dated examples.

The articles distinguish observations, technical explanations, deductions, opinions, and explicitly marked future hypotheses. A successful demo shows what is possible; it does not automatically show how reliably the result can be reproduced.

What counts as working?

Working code is only a minimum condition. Maintainability, changeability, testability, security, architectural fit, and economic appropriateness also matter. As code generation gets cheaper, the bottleneck may move toward requirements, decisions, architecture, review, and verification.

The same shift raises a human question. If agents economically replace many tasks previously used to train junior and mid-level developers, how will experienced senior engineers emerge? Junior work is both production and a learning path.

Articles

The LLM and Agentic Engineering glossary provides short definitions of the series’ core terms.