The Project
I administer our Datadog account and manage its costs. When the bill jumps, the explanation is almost never in Datadog itself. It is spread across what we released, what changed in our infrastructure, and what work was in flight. I built an AI agent to do that cross-referencing for me, and in time for my leads too, so we can get ahead of cost spikes instead of explaining them after the invoice arrives.
The problem
A cost increase starts as a number. Turning it into a cause means checking the latest releases, Jira tickets, change tickets, and merged pull requests, then lining them up against when usage moved. That is slow, manual detective work, and it tends to happen only after a surprise.
Deterministic data, AI analysis
The key design choice was to keep the model away from data gathering. The agent calls deterministic scripts that pull the evidence from each source, so it comes back the same way every time. The model’s job is limited to analysis: reading what the scripts return and reasoning about which change most likely explains the increase. It runs inside a coding harness and as standalone scripts.
Built with AI, directed by me
Most of the software was written by AI under my direction. I set the design and the boundaries, including the rule that the model analyzes evidence but never gathers it.
Wrap Up
It is still early, but it has already proved itself once: it traced a real Datadog cost increase to an infrastructure change. Next, my leads will use it so we can catch spikes as they start instead of after the bill.