Ideas
Short essays, notes and opinions on software engineering, published directly on elienehme.fr
2026
As soon as an action becomes recurrent, the agent can prepare it — but a system that speeds up the routine puts vigilance to sleep, while one that isolates the exception wakes it up
Conversation does not replace the gesture, it replaces the navigation that leads to the gesture: on high-frequency repetitive tasks, muscle memory is a feature
The rendering can be generative because the write is not: the guardrail is no longer a button hidden in an interface, it is a right evaluated whatever the wording
First use case, walked through end to end: same query, same data set, two forms of rendering, and no deployment between the two
Brownfield or greenfield predicts little: what matters is whether an agent finds the right context within clear boundaries, from day 1 to day 1,000
Whole production cycles were paid for to change a representation, while the data structure had not moved. It is not the screen that needs redoing, it is the idea that it must be frozen
Publisher, PowerBuilder, no-code: every tool was meant to end software development. Each took a task, never the craft: AI will too, if the foundations hold
Vibe coding lets anyone build a website, but it cannot make that website unique and safe
A vision is only worth what teams can follow: execution that is faithful and practicable, a pragmatic trajectory rather than fine ideas nobody can keep up with
Rebuilding generic skills that plugins and integrations already provide, and that the next models will absorb, is rarely the right investment
"AI will replace everyone", or, in the targeted version, "developers". Replacing a profession because its tool can now be generated confuses the two
A developer who only writes code will have less work, or none at all. The question is not whether AI writes the code, but what the developer brings beyond it
As AI writes more and more of the code, companies need to rethink where their sustainable competitive advantage will come from
Hexagonal architecture, clean architecture, CQRS: powerful patterns, often shown off without the good development practice that holds them up
A short list of what AI has changed in the work, a shorter one of what it has not, and what that means for the engineer’s job
2025
The model forgets everything between two requests: the chat application sends the conversation back each time. What users think the AI remembers is what the application chose to resend
A broker that delivers and forgets, a log that keeps and lets consumers read again: choose by what you need after delivery, not by throughput
No change should happen by chance, in the code or in an organisation. Five phases, from a pilot to a decision, turn a good practice into a standard that holds
A single expert per technology is a risk that looks like a strength. A shared calendar of rotating champions spreads the knowledge of each technology across the team and shows where the gaps are
A merge request approved by its author is a formality, not a review. The rule fits in one line of a team’s Git conventions, and it matters more as assistants write more of the code
2024
With OpenTelemetry, the standard belongs in what touches every service, its instrumentation and its collector, and the backend becomes a choice each team can change
Governing secrets takes two roles that work closely together and never merge: security writes the policy, operations runs the tool, and both choose it together
Putting everything on infrastructure as a service, or containerising everything as is with huge images, is not a migration: it is a move that costs more
2023
Code, test, describe the implementation: repeated in every user story, these tasks bring no business value, and Jira is not a development tool
Reviews happen at the wrong time, with the wrong question. What has made them useful in my experience
2022
A Kafka cluster per project looks like autonomy and costs like a debt: one shared offer, run by one central team, with every topic declared as code
With Kafka, receiving the same message twice is the normal case, not the exception: every consumer has to be written so that a second delivery changes nothing
Kafka is a distributed log with its own concepts: partitions, offsets, consumer groups, retention. Teams that bring their message-queue reflexes to it get it wrong, so the concepts come first
2021
The price of a technical decision is what it does to every decision that follows. Three questions to ask before making one
2020
Evolvability is measured by doing, not read off a diagram. Five unglamorous traits of codebases that age well