Beyond happy path engineering: Time

Beyond happy path engineering: Time
37 minutes by Gabor Koos

Gabor explains why time is more complicated in production systems than it first appears. He distinguishes wall-clock, monotonic, logical, and business time, showing when each should be used. He also covers clock drift, expiration, scheduled jobs, time zones, competing updates, and reconciliation. The main lesson is to treat time as a design boundary and make its assumptions explicit.

Ship long-horizon agents that hold up in production
sponsored by LangChain

Long-running agents break the assumptions behind standard app infrastructure. They can run for minutes or hours, pause for human input, recover from worker crashes, and still need controls around cost, memory, code execution, and tracing. Download the engineering guide to the runtime primitives that keep production agents reliable.

How to create your own decentralized messenger protocol
15 minutes by Michał Karbowiak

Michał guides on building a simple decentralized messaging protocol using federation. He covers server discovery, user identities, end-to-end encryption, public key exchange, message delivery, and server authentication. He shows how clients can encrypt messages while servers securely communicate and verify each other. The approach provides a practical foundation for building a simple, decentralized messenger without relying on a central server.

Notes on incidents
6 minutes by Sean Goedecke

Most incidents resolve without any action, and rushing to fix things often makes them worse. The best first step is to do nothing, take a breath, and observe. Simple fixes like disabling a feature flag usually solve the problem, but only if you know the system well enough to act decisively. Fixing incidents earns goodwill, but it is not a reliable path to lasting influence.

Growing as an engineer in a world of AI
18 minutes by Norberto Lopes

Norberto argues that leaning on AI to generate code without understanding it stunts your growth as an engineer. The learning that used to happen through struggle, repetition, and debugging is exactly what AI removes. To grow, he suggests treating AI as a thinking partner rather than an answer machine: ask why, type code out instead of pasting it, read the foundational books, and stay close to the hard conversations AI cannot join.

90% of the t distribution
3 minutes by Kqr

Gosset found that using a normal distribution to calculate confidence intervals gives results that are too narrow, because it treats an estimated standard deviation as exact. He developed correction factors based on sample size to fix this. With just two data points, you can estimate a standard deviation by multiplying the difference between them by 1.3, which helps judge whether a result is truly unusual.

And the most popular article from the last issue was:

• Harness engineering deep dive