Engineering Notes
Architectural decisions, performance constraints, and trade-offs from production: CNBC-scale edge systems, AI-agent engineering, and the discipline that holds them together.
-
A Judge Pays for Itself on One Kind of Sentence
Four note drafts went through seventeen independent judge passes this week, and two of them went live on this site. The judges raised fourteen blocking findings. Twelve were sentences that a dated record did not support:…
-
Merged Is Not Running: Three Ways a Landed Fix Stays Out of Force
A fix my fleet landed on a Sunday at 04:27Z was 33 hours from the first run that could execute it, and the prediction on file had treated the merge as the moment it took effect. The fix lived inside a scheduled heal in t…
-
A Self-Hosted Agent Fleet Inherits Its Host's Restart Window
At 04:59:01 on 2026-10-14 my Windows machine began restarting to install an update, one second after the last tool result of a scheduled agent run that had started at 04:10. The run belongs to an agent fleet I operate on…
-
Three Judges Cleared It, and All Three Had the Same Host
Three independent judge passes cleared a mechanism that was false on the machine the incident ran on. In September a generator in my workspace rewrote its output on every run of a Windows checkout, and the note explainin…
-
Fan-Out Multiplies Whatever Each Unit Pays to Start
Spreading work evenly across more units is the safe default, and it inverts the moment each unit pays a cost to start. Cloudflare documented it in September 2025. A visitor sending one request a minute into a data center…
-
Nobody Declared the Line Endings, and a Generator Never Converged
A code generator in my workspace judged its own output stale on every run and rewrote it. The repository carried no .gitattributes at all, so every text file took its line endings from whatever core.autocrlf the clone ha…
-
Two Checks Reported Zero and Neither Was Looking Where It Mattered
Two of my own checks spent this month returning a confident nothing. The tool that prices each scheduled run found no runs at all and said so. It carried two independent defects, either of which alone was enough: its pat…
-
Your Checks Are Lying to You
Run this in any repo with a gating script: ./check.sh | head -40; echo $?. If the gate fails, you still see 0, because without pipefail, $? is head's exit status, not the gate's. I measured it on a real gate in my own st…
-
Agents Re-Derive Judgment You Already Paid For
Every agent session starts from a blank context window, so it re-derives judgment you already paid for. You settle a tradeoff in March; in August an agent proposes the option you rejected, reasoning its way there with fu…
-
Reinforcement Anchors Beat Emphasis: Compressing a Production System Prompt
Production system prompts bloat by a predictable mechanism. The model does something wrong, so you add an instruction telling it not to. When that doesn't stick, you add a more forcefully worded one. NEVER do X. In capit…
-
The Agent Run Is the New Unit of Work, and Reviewing It Is Management
The genuinely new moment in AI-assisted engineering is not the chat answer; you watched that get produced and judged it in real time. It's when an agent comes back with finished work: it read the folder, edited the files…
-
Agents Degrade Quietly: Maintenance Is Where the Value Compounds
Building an agent produces a visible artifact (there was nothing, now there is a working agent), so it reads as progress. Maintaining one produces no artifact; at best, nothing happens. So effort flows to building, and t…
-
Spec Quality Is the Bottleneck Now, Not Implementation Speed
The industry is measuring AI-assisted development with the wrong unit of analysis. Code-generation speed is the vanity metric; the METR result everyone cites (experienced developers who felt 20% faster while measuring sl…
-
Agent Failures Are Loop Failures, Not Intelligence Failures
Every agent failure I've debugged this year decomposes the same way. The agent didn't lack intelligence. The loop lacked definition. It wandered out of scope because no boundary was stated. It "finished" without finishin…
-
Why I Made This Site Readable by Machines, Not Just Humans
A site now has two kinds of reader, and only one of them renders a page. Crawlers, retrieval pipelines, and agents pull the raw document straight into a context window and answer out of it, so a site built only for a per…
-
The Three Tiers of Using AI, and Why Only Two of Them Still Differentiate You
There's a real difference between using AI as a faster typist (autocomplete, chat-assisted edits, "fix this bug for me") and delegating a bounded unit of work to an agent that plans, executes across multiple files, and h…
-
Building with AI: The Compound Advantage
I built most of this site through Claude Code as a deliberate workflow: the components, the type errors that blocked deploys, and the notes themselves all went through the same loop of prompting and review. The obvious c…
-
The Front End Is a Distributed System, Starting with the URL
Filter a dashboard, refresh the page, and the filter is gone. Send the link to a colleague and they see a different view. The usual diagnosis is that the state was never persisted, and the usual fix is a client store, wh…
-
Decoupling State from Render in LLM Streaming
The naive way to build a streaming AI interface is to pipe a Server-Sent Events stream straight into a React state setter: a chunk arrives, setState(prev => prev + chunk) fires, the component re-renders. At sub-50ms toke…