Agent fleets in production: a field report
Running many coding agents at once works better than expected on one axis and worse on every other. What actually happens.
tech, developers, and the code underneath
older dispatches from the README archive.
Running many coding agents at once works better than expected on one axis and worse on every other. What actually happens.
Six causes, in order of how often they're the actual problem, with the fix for each.
Java, Python, C#, and JavaScript are all converging on the same feature from ML languages, thirty years late.
Running your own infrastructure got dramatically easier while the industry was arguing about the cloud. A practical assessment.
The average design doc is eleven pages and gets three comments, all on page one. Here's the format that works.
Phishing-resistant authentication is now the default on major platforms. The remaining problem is account recovery.
It shipped, it works, most of the web uses it, and almost nobody understands what changed. A practical review.
Not the hourly rate. The fragmentation. Here's why a 30-minute meeting costs four hours.
New silicon, new interconnect, and an industry where the unit of purchase is a room rather than a card.
Your code will be rewritten. Your API will be versioned. Your data model will outlive both and every mistake in it is permanent.
Frame budgets, determinism, and shipping to a fixed target. Game developers solved problems the rest of us are still arguing about.
Twenty flags is a million configurations. You are testing one of them. Here's how to keep that from being a problem.
Refusing work badly is a career problem. Refusing it well is one of the most valuable things a senior engineer does.
Container queries, :has(), view transitions, popover, anchor positioning. The polyfill era is over for real this time.
If your rotation is painful, that's information about your architecture, not about your people's resilience.
The capability floor rose faster than the ceiling. Most production inference no longer touches a frontier model.
Every team says docs matter. Almost none of them assign an owner, a budget, or a metric.
The fine-tuning wave, the RAG wave, and the agent wave all followed the same arc. Here's where the value actually settled.
Every type system decision is a trade between what you can prove and what you can express. Here's how to think about where to sit.
High-risk obligations arrive this summer. If you ship into Europe, the classification work should already be done.