System design, trade-offs and failure modes.
39 items
I modeled the KV-cache page reuse behind Cloudflare's GLM-5.3 leak. A fresh pool leaks nothing, a warm one hands over other customers' pages, and every reply still returns HTTP 200.
On 1,536 LoCoMo questions, tagging each fact with its subject lifted search from 54.4% to 59.5%. Filtering the raw turns by speaker did the same, for free.
On 1,536 LoCoMo questions, Hermes-style FTS5 session search found the answer turn in its top 10 for 54%, ahead of a small embedding model. Nearly every hit came from an OR retry.
The BIPM made UTC continuous from 20 May 2027, citing the risk of a negative leap second. I fed 23:59:60 to six runtimes, and none turned it into a moment of its own.
Anthropic replaced git for its monorepo. I measured what shallow, partial and sparse checkouts do to a 343,502-commit repository, and where stock git still runs out.
DEDA decodes the hidden tracking dots laser printers stamp on pages. I generated my own pattern, decoded it, then corrupted one dot and watched the checksum fail while the decoder answered anyway.
Red Hat benchmarked Jev against 8 other guardrail models and found a local classifier beats it on speed by 6x. I tried to reproduce that result myself, on my own task.
liao.gg argued vector-based agent memory cannot tell correct from stale, without a benchmark. I tested the claim on this blog's own corrected numbers, with real embeddings.
Cloudflare launched an OHTTP gateway that hides IPs from app servers. The RFC and Cloudflare both say privacy depends on relay and gateway never sharing an operator, a rule no cryptography enforces.
Git is replacing SHA-1 with SHA-256 for every commit and file it tracks. A GitHub co-founder argues the attacks motivating the switch do not threaten how Git actually works.
Netlify replaced the lightweight sandbox running its edge functions with full virtual machines, and median latency dropped from about 30ms to 5 to 6ms.
An open-source benchmark set out to catch Claude quietly getting worse after launch. Its own validation test, swapping in a different model, came back statistically indistinguishable from noise.
The same classification task, run through Jev and Gemini 3.8 Flash with real API calls: identical accuracy, four to five times the latency, eight to thirty-four times the cost.
Ten processes update the same git ref at once. Git returns success for all ten, and by default it hides which nine of them never actually stuck.
DeepSeek's DSec runs 380,000 concurrent AI-agent sandboxes a day. Not by caching one big image, but by admitting most of a sandbox can never be reused, and building around that instead.
Go 1.27 ships a portable SIMD API with zero published benchmarks. Measured on real hardware: a plain loop wins below 256 elements, then SIMD holds a steady ~2x, not the 8x its lane count implies.
Gambit Security recovered the staging server behind a three-agent breach of 27 companies. The same ungated pipeline destroyed its own stolen data — the exact failure mode this blog already mapped.
Docker Sandboxes isolates AI agents in microVMs, not containers. Two bugs patched in the same release both broke it at the one place the guest is allowed to talk back to the host.
Three companies shipped a standalone agent browser. One matched Windows to macOS day one. One took 83 days. One never got there, shut down, and retreated into an extension.
I gave one reviewer agent nothing and another a 210-word handoff note. Cognition says blind works better. The one with the note caught the bug that actually mattered.
Two subagents build the two halves of a receipt module in total isolation. One half won't compile. The fixed version compiles, runs, looks correct — and is silently wrong by a cent.
Two subagents with identical toolboxes and different system prompts boot 104 tokens apart. Add four tools and the gap is 17,252. The boot bill is the toolbox, and two tools are most of it.
A subagent that runs one wc -w burns 34,172 tokens. Split the work across five and the bill is 4.85x for a 1.79x speedup - the multiplier is a boot sequence, paid again every time you fan out.
Container isolation for a fleet of agents costs 48 seconds once and 4.1 kB per agent after — until the bind mount that makes it convenient costs 465 MB each and hands the sharing back.
An ADE gives every agent its own git worktree. Here that isolates 44 files and leaves 23,199 shared — and the usual fix is a symlink, which is the opposite of isolation.
The web's oldest isolation guarantee is a string comparison. Put an agent inside the browser and it becomes a model's ability to resist a prompt injection.
A multi-turn agent replays its entire history on every request. Caching, context editing and compaction each attack a different part of that bill — and one of them, at its defaults, made it worse.
Hermes Agent says it improves its own skills. In the source, that loop is a tool-call counter, a forked LLM review told to be active, and a curator that archives what nobody uses.
Rendering is a separate, queued, stateless stage between crawling and indexing. Almost everything confusing about JavaScript and search comes from ignoring that.
Prefill and decode have opposite performance characteristics, and almost everything confusing about LLM latency, cost and context limits falls out of that one split.
ego lite and Aside are both browsers built for AI agents, but one has no agent of its own and the other ships with one. A guide to what each is, with the same page-reading test run on both.
A BFF gives each client its own server-side backend. This guide covers the problem it solves, a runnable demo with measured request counts and bytes, its security role, and when to skip it.
Augmented engineering has no standard definition, while agentic engineering does. This guide sorts out the usages, compares the two working styles side by side, and says when each one fits.
Claude Duet is Anthropic's proactive personal agent. This guide explains what it does, how it asks for approval, and how it differs from Meta Muse and OpenAI dots, using the vendors' own pages.
A beginner-friendly SRE guide to service level indicators, objectives and agreements, with a comparison table, the downtime each "nine" allows, and a real cloud SLA read line by line.
A beginner-friendly comparison of RDF triples (SPARQL, OWL) and property graphs (Cypher), checked on one fact, one class hierarchy and a 1M-edge graph, with runnable code.
Baalda is a new open-source notes app that also keeps plain Markdown files on your disk. A source-checked comparison with Obsidian on collaboration, AI access, price, plugins and maturity.
How gRPC decides which backend gets each call, why a TCP-level load balancer leaves servers idle, and four ways to fix it. Every distribution number comes from a local experiment you can rerun.
Four ways to build an API, compared on who decides the response shape, what the contract looks like, how errors and caching work, and how big the bytes are, with a size measurement you can rerun.