A render farm knows how to handle a crash.
It has no idea what to do with a frame that worked.
A texture fails to resolve on the worker. The renderer substitutes a placeholder, prints a warning, and exits 0. The frame saves. The frame count is complete. The durations look normal. Every automated system in the pipeline calls it a success, and the jacket comes back flat magenta.
Nobody finds out until a human watches the dailies. This is the failure Dailies catches, from telemetry, before anyone watches anything.
Not a hypothetical
The material will fail to load silently, rendering as black. The render completes without error messages, but the output is incorrect.
Even if only one texture is missing, the render will submit a 'done' status. You discover it after the render completes, and you have already paid for those broken frames.
Textures are a silent cause of instability.
Live, from the farm, right now
Real shots on the deployed system, read from Prometheus as this page loaded. Nothing here is seeded or illustrated.
The case that matters is a shot whose frames are all complete and which is still wrong. Every number a scheduler looks at says success. The evidence is a log line and the picture itself, which is why the agent reads logs as well as metrics, then opens the frame. Every finding it reports names the query behind it, and every visual verdict names the frame it judged.
Open the board and press DiagnoseHow it works
Blender renders
A real render runs as a Cloud Run job. Its stdout is parsed into typed events by the same parser the tests cover heaviest.
OpenTelemetry ships it
Frame durations, memory and job progress go to Prometheus. Render-domain failures go to Loki, because a missing asset is a sentence, not a number.
Grafana holds the truth
The board reconstructs every shot from those series. There is no database of shots to drift from the farm, and nothing to seed.
The agent investigates
A Gemini agent on the Agent Development Kit reaches Prometheus and Loki only through the Grafana MCP server. Remove that connection and the project stops rather than degrades.
And then it looks at the frame
Gemini opens the picture the render actually produced. It is told how renderers signal a failed texture and nothing at all about what this shot's metrics or logs said, so it can disagree with them. Two sources that cannot disagree are one source.
What one diagnosis costs
The agent is measured the way the farm is. If you cannot trust what you cannot see, that has to apply to the thing doing the looking.
- Wall clock
- 40.3s
- Model turns
- 7
- Input tokens
- 27,313
- Thinking tokens
- 5,124
- Output tokens
- 878
Measured on a real investigation, in OpenTelemetry GenAI conventions, in the same Grafana stack the renders report to. Thinking is counted separately rather than folded into output: it is billed, it appears in neither the prompt nor the answer, and it is six times the size of the reply. Folded in, this would look like an 878-token call.