Share with your CTO
GitHub is giving engineering leaders what they’ve needed since agent apps became a real workflow layer: per-agent visibility inside the Copilot usage metrics API. The update adds a totals_by_3rd_party_agent array that breaks out activity by individual agent, covering partners like Claude and Codex. Each entry reports user-initiated job starts and session counts across 1-day and 28-day windows, available at both the enterprise and per-user level. Before this, all agent activity collapsed into a single bucket, making comparative adoption analysis impossible.
What this means for your business
The immediate problem this solves is agent sprawl opacity. Engineering orgs that rolled out multiple agents simultaneously had no way to know whether Codex was displacing Claude adoption or whether both were stalling at the same team boundary. Now you can see which agents are actually pulling adoption and which ones are ghost licenses, identified by stable agent_id rather than display name. That’s the difference between a rollout decision grounded in evidence and one grounded in whoever talked loudest in the last planning meeting.
There’s a sharper implication here. The metrics API becoming agent-aware signals that GitHub is positioning itself as the control plane for multi-agent developer workflows, not just a host for individual tools. When the observability layer covers third-party agents, the platform that owns that observability owns the procurement conversation too. A CTO who standardizes on GitHub’s metrics pipeline for agent reporting is implicitly anchoring their entire multi-agent stack to GitHub’s taxonomy of what counts as an agent, how activity is counted, and what gets reported upward.
The tradeoff is real: you gain actionable adoption data now, but you’re building your agent governance reporting on top of a vendor-defined schema. If GitHub changes how it classifies or collapses agent activity, your historical comparisons can shift under you. The signal worth watching is whether GitHub opens that schema to community input or treats it as proprietary infrastructure.
Concept deep-dive: Agent app sessions vs. user-initiated interactions
The API distinguishes two counts that look identical but measure different things. A user-initiated interaction counts discrete job starts, each time a developer explicitly kicks off an agent task. A session groups related activity into a continuous working period, similar to how a web analytics session groups page views by time proximity. The critical warning in the release: these two fields share a name at different levels of the JSON response and must not be summed together. Confusing them would inflate your usage numbers and distort any ROI calculation built on top of the API.
Based on reporting from Copilot usage metrics API adds agent app activity, originally published 2026-08-07 14:20:00.
