I build AI agents and detection systems for email security at Abnormal AI. Its platform operates at billion-message scale. I write about detection engineering, reliable agent execution and the systems behind them.
Technical notes, longer guides and occasional essays. Follow via RSS or Substack.
An online migration can quietly lose its rollback
The first write that reaches only the new store changes your rollback procedure. Reverting the application might still work at the deployment layer. If the old code reads the old store, the user can get an answer with yesterday’s state.
That is an easy trap to miss in a migration plan that says “compare, switch, roll back if needed.” Put an expiration condition on the rollback. You want to know which write ends it, who can stop that write path, and what recovery looks like after it happens.
...
A canary needs enough traffic to fail
If 1,000 canary requests pass, I’d still ask what failure the test had a realistic chance of seeing. If the defect hits one in 10,000 requests, a clean run of that size is the likely outcome even with the defect present.
A rollout plan usually starts with an exposure limit. I want a second number on it, the minimum amount of relevant traffic I need to see before the next decision. The two numbers move against each other when the canary is small.
...
An agent benchmark needs a resource budget
A six-point lead on a coding benchmark would get my attention. So would learning that the same agent gained six points when its test environment stopped killing processes at a strict resource limit.
Anthropic reported that gap between its tightest and uncapped Terminal-Bench 2.0 configurations. The model comparison people want to make is harder than it looks. The agent is writing files, running tests and pulling dependencies inside a real environment. A change to the worker can change the work it manages to finish. Anthropic’s infrastructure experiment · Terminal-Bench 2.0 release
...
How the AI buildout reaches your DDR5 bill
Part of The AI Infrastructure Reality Check.
On November 20, Corsair wanted $643.99 for a 32 GB kit it couldn’t ship. Two months earlier, a snapshot showed the same kit in stock at $559.99. The $84 difference is easy to calculate. What a buyer actually paid is still an open question.
The trail gets clearer at the factory. Micron says it moved some DRAM supply to data-center and cloud customers in response to AI demand, with an emphasis on high-bandwidth memory. In November 2026, TrendForce reported slow PC sales and rising inventories of finished PCs. It still expected PC DRAM contract prices to rise 13–18% quarter over quarter in Q4. PC makers can sell fewer computers and pay more for memory if suppliers cut allocations faster than orders fall. Micron FY2025 Form 10-K · TrendForce November PC procurement report
...
Six AI businesses and five tests of the boom
Part of The AI Infrastructure Reality Check.
NVIDIA reported $107.6 billion in Data Center revenue for the quarter ended October 25, 2026. For the quarter ended September 30, CoreWeave reported $3.66 billion in revenue and a $709 million net loss. Those totals cover different periods and different businesses. The chip supplier gets paid for hardware. The cloud operator needs to rent out capacity long enough to cover what it cost to put into service. NVIDIA Q3 FY2027 results · CoreWeave Q3 2026 results
...
GPU economics from break-even to cost per token
Part of The AI Infrastructure Reality Check.
In the model below, a GPU earns $2 an hour for 90% of the year and still finishes $1,212 short. The operator bought a share of the server and network along with the GPU and needs to recover that investment over four years. At $2.50 an hour, the same installation breaks even at about 77% billable time. The GPU break-even calculator lets you price your own assumptions.
...
What limits an AI cluster from delivery to peak load
Part of The AI Infrastructure Reality Check.
In the model below, the GPUs arrive in month 12. The operator can send its first bill in month 25. Move the delivery up four months and the invoice date stays put. Finish the grid work four months early and billing moves up one month. The delivery date on a purchase order tells you when the equipment gets to the site. Billing depends on when the operator has a working cluster the customer can accept.
...
AI infrastructure cash flow and who finances the buildout
Part of The AI Infrastructure Reality Check.
From the September 2025 quarter to the same quarter in 2026, Microsoft, Alphabet and Meta brought in $9.6 billion more operating cash. They spent $62.5 billion more on property and equipment. After those purchases, the three companies had $8.41 billion left between them, down from $61.29 billion a year earlier. Microsoft Form 10-Q · Alphabet results · Meta results
The bridge compares the same September quarter in each year. It uses consolidated cash flows and excludes the separate finance-lease adjustment.
...
You Cannot Grade an Agent You Only Read
I wanted to check whether a skill actually did its job. The skill ran a goal prompt through OpenCode, called a handful of tools, and left a session behind. Then I read the final message, and the final message told me exactly what I wanted to hear.
That is the problem with grading an agent from its output alone. The reply is the last thing that happened, not the thing you were measuring.
...
The Model Is Not the Agent — Inside DeepSeek Harness
Also available on Archit’s Newsletter on Substack.
You ask a coding agent to fix a failing test. Somewhere, a model generates tokens. But tokens do not open files, decide which shell command is permitted, preserve a session after a crash, or establish that the test now passes. The execution system has to make those things happen.
That system is the harness. Its decisions determine how reasoning becomes action, what the model sees next, and which failures remain recoverable. To understand an agent’s reliability, we need to follow those boundaries in code.
...