Kenya's tech careers, reported dailyWhatsApp channel · info@thetechhut.co
The Tech Hut.
Lines of colourful code on a dark editor screen
The Reading List · Nº 4

5 things worth reading this week

The Tech Hut ·

This week: how GitHub rebuilt Git storage so thousands of AI agents can push to one repository; keeping AI-written code honest with tests written as data; durable execution for agents that run for days; the unusual consensus algorithm Cloudflare picked over Raft; and what an attack using paid human CAPTCHA solvers looks like from the server's side.

Infrastructure

Inside GitHub's new Git storage architecture

Karthik Raman and Brian Celenza · The GitHub Blog · 5 November 2026

Busy repositories used to see a few dozen writes a second in short bursts. With AI agents pushing around the clock, one repository's daily requests quadrupled in two months, so GitHub rebuilt how Git data is stored. Machines that serve a repository now keep only a cache on disk; Git objects live in cloud blob storage, and only the branch and tag pointers go through a small replicated log that commits each update by quorum. In testing it handled 5,000 agents on one repository and 7,200 pushes a second. It is a clean lesson in separating the small piece that needs agreement from the large piece that does not.

Testing

Tests as data: ensuring AI-generated code does what we want

Matteo Vaccari · martinfowler.com · 11 November 2026

When an AI agent writes both the code and its tests, it can quietly weaken the tests to make them pass. Vaccari's answer is to write the important tests as a plain table, in YAML or CSV, with the name, input and expected result on each row, and let one small generic runner execute them. A person can read the table in a minute, see which cases are missing, and notice if a row changes. He extends the idea to stateful features with a given-when-then text format whose failures show up as readable diffs. A practical technique you can try on your next project, with or without AI.

AI

Long-horizon agents and durable execution

Kiran Prakash · martinfowler.com · 10 November 2026

An agent that reorders stock or chases a supplier may run for days and will meet crashes, timeouts and odd replies. Durable execution runtimes such as Temporal, Restate, Inngest, DBOS and Cloudflare Workflows record the result of every call to the outside world, so a crashed workflow can replay and carry on where it stopped. Prakash's point is that this is necessary but not enough: you still need idempotency keys so retries do not place two orders, compensating steps to undo work, and checks that a decision made before a long wait is still valid. Read it before you build an agent that does more than chat.

Distributed systems

Choosing QuePaxa, an unusual consensus algorithm

João Pedro Leite, Bob Halley and James Larisch · The Cloudflare Blog · 11 November 2026

Raft and Paxos rely on one leader, and when that leader slows down the whole system can stall; Cloudflare once had a dashboard run 80 times slower for six hours because of a failing switch. This post explains QuePaxa, a less familiar algorithm in which any node can make progress if the leader fails, and why Cloudflare chose it for a new control-plane service. In their tests, decisions fell to about 65% of normal during a leader failure and recovered within a minute, without ever stopping. The authors are candid that it is less battle-tested than Raft. A rewarding read if you want to understand consensus beyond the textbook.

Security

What a CAPTCHA Farm Attack Actually Looks Like

Reetain Raina · freeCodeCamp · 4 November 2026

A CAPTCHA does not stop a bot that pays people to solve it: solving services charge roughly half a dollar to two dollars per thousand puzzles and answer in 10 to 25 seconds. This article shows the traces such an attack leaves, such as forms filled instantly followed by a pause, no natural mouse movement, data-centre network fingerprints and pages visited without loading their images, and the defences that work: watching the whole session rather than the puzzle, fingerprinting, proof-of-work rate limits and monitoring after login. Useful if you run sign-ups, ticketing or anything bots like to farm.

How we pick: each week we go through the developer feeds we follow, read the originals, and keep the few a developer in Kenya would want to know about. Every link goes to the author’s own page.

The Reading List is a weekly set of developer articles picked by The Tech Hut. Every link goes to the original author; the notes are ours.