Video Game Data Pipeline

A leading game studio needed to turn raw gameplay recording sessions into clean, delivery-ready annotated data at a pace manual review couldn't sustain. We built an end-to-end annotation pipeline for video game data collection, anchored by 4 agentic workflows: automated QC validation, delivery preparation, cloud sync, and real-time notifications.

Game recording annotation has a throughput problem that's structural, not incidental. Sessions arrive continuously, each one needs the same battery of QC checks before it can ship, and manual review of every session simply doesn't scale to the volume a studio's data collection program generates. The studio needed a pipeline where quality gates, packaging, and delivery all happened automatically — with humans reviewing only what the pipeline flagged, not everything that came in.

The bottleneck was never capturing the gameplay — it was everything that had to happen to a session between capture and a model-ready dataset.

Rather than one monolithic script, the pipeline is built from 4 agents, each owning one stage:

Delivery had to work across the studio's existing cloud footprint, not force a migration to a single provider. The cloud-sync agent authenticates over OAuth and syncs packaged deliverables to both GCS and R2, so the studio's downstream tooling — wherever it lives — can pull from the storage backend it already integrates with. Delivery-ready packages are produced automatically by the delivery-preparation agent immediately after a session clears QC, with no manual repackaging step in between.

A pipeline that runs unattended still needs to be observable — a session silently stuck between QC and delivery is as costly as one that fails outright. The notification agent closes that loop: the moment a session passes QC, fails validation, or completes delivery, the responsible team is alerted in real time rather than discovering the state days later during a manual audit. That real-time signal is what let the studio trust the automation enough to stop checking every session by hand in the first place — the pipeline surfaces exceptions on its own, instead of asking a human to go looking for them.

An automated pipeline nobody can see into just moves the manual-review bottleneck downstream. Real-time notification is what makes the automation trustworthy enough to actually rely on.

The studio now runs game recording annotation on a pipeline built from 4 agentic workflows — automated QC, delivery preparation, cloud sync to both GCS and R2 over OAuth, and real-time tracking of every session's status. QC and delivery packaging that used to require manual, session-by-session attention now run automatically, with the team's time going to the sessions the pipeline actually flags rather than every session that comes in.