Episodes

Episodes represent complete video productions that move through an 8-stage content pipeline, fully orchestrated by AI agents with automated quality gates.

Pipeline Stages

Each episode progresses through these stages sequentially. The worker-orchestrate service coordinates execution, invoking specialized agents and managing feedback loops automatically.

Stage	Agent	Reviewer	Max Iterations	Description
Research	Researcher	Producer	2	Gather facts, sources, competitive analysis, define unique angle
Script	Writer	Producer	3	Write screenplay in channel character voice
Storyboard	Storyboarder	Producer	2	Plan visual sequences and scene layouts
Assets	Asset Finder	—	1	Find and download media resources (stock footage, images, audio)
Compositing	Compositor	Producer	3	Assemble Remotion video composition from 65+ components
Export	Render Worker	—	1	Render final MP4 via Chromium (dispatched to worker-render)
Review	Producer	—	1	Holistic quality gate across all artifacts — always requires human approval
Publish	—	—	1	Guard on review approval, mark episode complete

Execution Modes

• Full Pipeline — click "Run Pipeline" to execute from the next pending stage through to completion (or until a stage requires human intervention)
• Single Stage — click "Run" on any individual stage to execute just that stage

Quality Gates

The producer agent scores each stage output 1-10. Stages scoring 7+ pass automatically. Failed stages trigger a feedback loop:

1. Producer writes structured revision notes
2. Notes are saved as stage feedback
3. The responsible agent re-runs with the feedback context
4. This repeats up to the stage's max iterations

If max iterations are exhausted without passing, the stage is marked needs_human and the pipeline pauses for manual intervention.

Upstream Issues

If the producer identifies a problem originating from an earlier stage (e.g., a weak script caused by thin research), both the current stage and the upstream stage are marked changes_requested. The pipeline halts so a human can decide whether to re-run the upstream stage.

Soft Recovery

If an agent hits its turn limit but still produced the expected artifact (e.g., research.md exists), the orchestrator proceeds to review rather than hard-failing — the artifact may be good enough to pass.

Stage Statuses

Status	Meaning	Color
not_started	Stage hasn't been run	Gray
in_progress	Agent or render currently executing	Blue (pulsing)
review	Producer is evaluating the output	Yellow
complete	Stage passed quality gate	Green
approved	Human explicitly approved (review stage)	Green
changes_requested	Feedback provided, needs re-run	Gray
needs_human	Max iterations exhausted or human gate	Pink
error	Agent or render failed	Red

Episode Files

Each episode lives in channels/{channel}/episodes/{slug}/ with:

manifest.yml          # Pipeline state (stage statuses)
research.md           # Research output
script.md             # Written screenplay
storyboard.md         # Visual plan
Episode.tsx           # Remotion composition
feedback/             # Producer review notes per stage
  research.md
  script.md
  ...
output/               # Rendered video
  {slug}.mp4

TIP

Episodes require NATS workers to be running. Workers are deployed as Kubernetes Deployments in Helm (see the workers values section). For local dev: docker compose -f docker-compose.dev.yml up -d.