Layer two: orchestration. The layer that gives the orders. This is the layer that lists the steps, runs them in order, retries what fails, and shows exactly where something broke.

01 Interfacewhat you click
02 Orchestrationwhat gives the orders
03 Computewhat does the work
04 AI runtimewhat makes the judgment calls
05 Statewhere everything rests
Five layers under every AI system. We go top to bottom: the surface you touch, down to the bedrock everything rests on.
What gives the orders. Workflows that retry, wait for free, and show exactly what broke.
What gives the orders. Workflows that retry, wait for free, and show exactly what broke.

The recipe the computer follows

The moment a file lands in storage, an event fires, and a Step Functions workflow wakes up. Think of it as a recipe the computer follows, step by step. It lists every step, runs them in order, retries what fails, and shows exactly where something broke.

1 S3 eventthe upload trigger wakes the workflow
File sanitizerchecks codec, container, duration
3 Already processed?a choice that prevents double work
Preprocesshands off to MediaConvert
5 Labels, transcript, meaningthree analyses fan out in parallel
6 Merge to briefthe streams come back together
Still from the module video

Nine workflows, and the main one only delegates

VideoFlow runs nine of these workflows. The main one drives the whole pipeline, and it never does any work itself. It only delegates. The orchestrator is a conductor, not a worker. It decides what happens next and hands each task to the layer below.

The zero-compute poll

Here is the pattern that makes it cheap. To wait on a slow job, the workflow asks: are you done? If not, it pauses. Completely.

Nothing runs during a Wait state. Polling a transcription job costs zero compute.

One pattern, reused for Transcribe, Bedrock Data Automation, and MediaConvert. Every slow job in the system waits the same free way.

Orchestration is also where time lives

EventBridge fires the schedules, queues decouple the work, alarms watch the metrics.

EventBridgefires the schedules
2 Queuesdecouple the work
Alarmswatch the metrics

When it fails at 2 a.m., no log hunting

And when something fails at 2 a.m., the execution history shows every state with its input, its output, and the exact error, inline. No log hunting. You do not go digging through scattered logs to find the failure. The failed state is right there, with what went in, what came out, and the exact error.

An assembly line where every station is visible, and the conveyor belt itself is free.

But orders are nothing without workers. Layer three.