The server I needed cost half the quote

On 1 November I asked an agent how big a remote server I would need to run my agent sessions there instead of on my Mac. It came back with a confident answer: a 16 GB machine with 4 CPUs and 320 GB of disk, at $84 a month, plus roughly $5 to $15 a month for backups.

I can’t judge a number like that. 16 GB sounds reasonable. It’s also the amount of memory my Mac already has. So I asked the agent how confident it really was.

Figure: The same day produced a quote, a measurement and a smaller recommendation.

What the first answer was built on

The price was real. The agent had read the cloud provider’s public pricing page for AWS Lightsail: $84 a month for the 16 GB bundle, and snapshots at $0.05 per gigabyte-month.

The size was not measured. The agent had assumed I would run 3 to 6 agents at a normal time, 8 to 12 light ones at a peak, and one or two heavy build or browser jobs at once. Those counts came from its guess about how I work. Nobody had counted anything.

The measurement

After I pushed, the agent looked at my Mac. It ran a read-only check of running processes. It read sizes only, no command arguments and no credentials.

It found 22 processes belonging to my agent app. Added up, they held about 1.70 GiB of memory. That included the app window, its server, one agent runtime and the helper tools. One session was actually running.

The new advice called 16 GB a comfort option, not a minimum. Start with the $44 machine, keep heavy jobs in a queue, watch the real peaks, and move up only if memory gets tight. The $84 tier stays as the upgrade path.

The gap is $40 a month, or $480 a year, for a server I hadn’t yet shown I needed.

What the measurement doesn’t prove

The agent was careful about this part, and I want to be as well. The number is one snapshot on a Mac, with one active session. It isn’t a peak. It isn’t several agents at once. It isn’t a Linux server. Adding up memory per process can also count shared memory more than once, so 1.70 GiB is a rough figure.

I can’t say 8 GB is enough. I can say 16 GB was a guess presented as a requirement, and the guess had a price on it.

What I do now

When an agent recommends a size, a count or a budget, I ask for the measurement behind it. If the answer is “an estimate”, I ask whether something can be measured in five minutes, read-only, before I spend money.

Before I spend money on this, show me what it's based on.

1. For every size, count or price in your recommendation, say whether you measured it or estimated it.
2. For each estimate, can you measure the real number in five minutes without changing anything? If yes, do that now, read-only.
3. Then give me the recommendation again from the measured numbers, with the cost, and the smallest option that would work.

Here that took one question. The first answer sounded sure and was built on assumptions. The second was built on a number from my own machine, and it was half the price.

I’d rather pay for a pilot and a measurement than for a quote.