If you cannot name the agent, you cannot manage the AI budget
Teams need per-agent records for identity, user, session, task, cost, outcome, and approval state before they can govern AI spend.
TL;DR
AI budgets become guesswork when every agent lands in one usage total.
GitHub has started separating recognised third-party agent apps in its Copilot usage metrics. The new data identifies the agent, its sessions, and user-initiated activity. That is a useful step because an organisation cannot judge agent cost or value if it cannot say which agent did the work.
A practical agent ledger needs seven fields: identity, user, session, task, cost, outcome, and approval state.
What changed
GitHub’s Copilot usage metrics update adds activity from recognised third-party coding agents. GitHub says the API can now report stable agent identifiers, sessions, and user-initiated starts.
This does not provide a complete cost or effectiveness system. It does separate agent activity that was previously harder to distinguish inside a blended Copilot total.
That distinction matters as teams add several agents to the same workflow. A coding agent, review agent, support agent, and internal operations agent may all create activity. Their operating costs and useful outcomes are different. One total hides those differences.
The seven-field agent ledger
Start with one row per run or bounded task:
| Field | Record | Question it answers |
|---|---|---|
| Agent identity | Stable name and ID | Which agent did the work? |
| User | Person or service that started it | Who requested or owns the run? |
| Session | Run or conversation ID | Can we reconstruct the work? |
| Task | Short, bounded description | What was the agent meant to do? |
| Cost | Model, tokens, tools, and estimated spend | What did the run consume? |
| Outcome | Completed, partial, failed, or abandoned | Did it produce anything useful? |
| Approval state | Automatic, pre-approved, reviewed, or rejected | Which human control applied? |
The first three fields give attribution. The remaining four connect activity to business value and control.
Why token totals are not enough
Token usage can show consumption. It cannot show whether the agent solved the problem, duplicated another run, waited on a broken tool, or produced work that a human rejected.
The same applies to session counts. More sessions might mean adoption. They might also mean retries, routing loops, or a workflow that keeps losing state.
Useful telemetry connects spend to an outcome:
- Which agent completed the task?
- How many attempts did it need?
- Did a person approve the result?
- Was the work accepted, revised, or discarded?
- Could a cheaper agent or model have handled it?
Without those answers, an AI budget measures activity rather than performance.
A weekly operator review
A small team does not need a large observability platform to begin. Give one person ownership of a weekly review and ask:
- Which agents created the most completed work?
- Which agents consumed the most time or money?
- Where did retries, timeouts, or abandoned sessions cluster?
- Which work required human correction or approval?
- What should be stopped, rerouted, or tested next week?
The review owner matters. A dashboard without a decision routine becomes wallpaper.
Rob’s take
The next AI cost problem will not be a single expensive model call. It will be a collection of agents producing activity that nobody can attribute to a useful outcome.
Per-agent telemetry turns that fog into an operating record. It will not prove an agent is effective on its own. It gives the operator enough evidence to ask the right question: did this agent earn what it consumed?
Watch next
- Cost and latency broken down by stable agent identity
- Outcome tracking linked to sessions and tasks
- Approval records attached to agent runs
- Budget controls that can stop or reroute one agent without disabling the rest
Source
- GitHub Copilot usage metrics API adds agent app activity, 7 August 2026
Quick signal helps Rob sharpen future briefings.