The next agent fight is going to be full of spreadsheets.
GitHub’s June 19 Copilot changelog adds a small field with a very large shadow: ai_credits_used now appears in the Copilot usage metrics API at the user level. Enterprise administrators and organization owners can see how many AI credits each user consumed per day, including single-day and 28-day reports.
This is not flashy. That is why it matters.
The invoice needs a map
Agents turn software usage into something much harder to predict than seats.
One developer asks a short question. Another asks an agent to inspect a repo, plan changes, run tools, edit files, rerun checks, and open a pull request. Both are “using Copilot,” but they are not creating the same cost footprint.
GitHub is not saying this new field is the invoice. It explicitly says the metric is a consumption signal, not a billed total, and that it is not currently broken down by feature, model, or surface.
Still, putting credit burn next to usage is the right primitive.
Companies need to answer boring questions before agent adoption gets serious. Which teams are actually using the system? Where is spend rising? Is the usage tied to valuable work, or just expensive curiosity? Are a few heavy users creating the whole bill? Does the team need training, model routing, policy, or just a bigger budget?
You cannot manage any of that from vibes.
Agent FinOps is becoming product infrastructure
This fits the pattern from the last few weeks. Microsoft is talking about agent cost control. GitHub is adding usage metrics. Enterprise tools are starting to expose the operational layer around AI, not just the magic box in the editor.
That is where the market is going.
The winning developer agent products will not only be better at code. They will make cost visible, policies enforceable, model routing explainable, and agent work auditable enough that managers do not panic every time someone lets the tool run.
There is a very human reason for that.
People love autonomy right up until the bill arrives without a story attached.
If agents are going to become normal infrastructure, the credit meter has to move from the finance department into the workflow itself. GitHub’s change is early and incomplete, but it points in the right direction: every serious agent platform needs to show what the work consumed, who triggered it, and whether the output justified the burn.
Source: GitHub