Blog category
Articles on AI-Assisted Engineering from our project work
Everything devs group has written on AI-Assisted Engineering, newest first.
AI assistants multiply the amount of code a team can produce. They do not multiply the amount of code a team can understand. The technical debt of the next few years is being created in that gap, and that is exactly where this category starts: how do you use coding agents so that there is less code at the end, not more?
AI-Assisted Engineering
The agents nobody approved
OpenClaw runs locally, without procurement, with far-reaching rights. Agent governance means: an allowlist, permission classes, detection.
AI-Assisted Engineering
Guardrails for coding agents: never its own reviewer
Scale review depth by task type rather than by agent confidence. What belongs in the CI pipeline and what a human has to read.
AI-Assisted Engineering
Using coding agents in line with data protection law: the gap in the subscription
A personal Pro subscription used on company code falls under consumer terms. Which plans and hosting routes make coding agents compliant.
AI-Assisted Engineering
AI productivity: METR's 19 per cent is out of date
METR's follow-up study flips the sign. What the research actually shows in August 2026, and why both camps quote it wrong.
AI-Assisted Engineering
AI writes the code. Who writes the architecture?
How we use AI-assisted engineering so that less code exists at the end, not more.
What this category covers
The articles cover the whole chain: the architectural responsibility that stays with a human, guardrails for coding agents and the rule that an agent must never be its own reviewer, the data protection question when coding agents are used inside a company, and a sober look at the productivity studies from METR through 2026, whose most-quoted number is out of date.
The standard never changes: whatever an agent writes goes through the same reviews, tests and architecture decisions as human code. What that looks like in practice we show in projects and in a first conversation; the groundwork is on the artificial intelligence and machine learning page.
The articles rest on sourced evidence, from the METR studies to the vendors' own documentation, and on daily work with coding agents in client projects. Where the research is thin or contradictory, the text says so rather than promoting a single number to the status of truth.
This field moves faster than any other category on this blog, so the articles carry visible dates and statements that have been overtaken get corrected rather than left standing. To follow where things stand, subscribe to the RSS feed or check the blog index.
For teams introducing coding agents right now, the order of the articles is almost a plan: settle the architectural responsibility first, then the guardrails, then data protection, and calibrate your productivity expectations against the studies rather than the keynotes. Where a team wants company along the way, we take on the setup, the review processes and the training; the first conversation about it commits you to nothing.
What this category deliberately does not contain is a debate about whether AI assistants are changing software engineering. They are, in our projects, every day. The interesting question is how, and it is decided by unglamorous things: review processes, permissions, the discipline of putting architecture before generation. Those unglamorous things are what the articles are about.