AI Agents
CRAFT uses 14 specialized AI agents that collaborate through the content pipeline. Each agent has a defined role, structured output format, and quality expectations. Agents are invoked via the Claude Agent SDK with full tool access (Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch).
Pipeline Agents
These agents are directly wired into the 8-stage orchestration pipeline:
| Agent | Pipeline Stage | Tools | Output |
|---|---|---|---|
| Researcher | Research | Web search, file read/write | research.md — sources, data, competitive analysis, unique angle |
| Writer | Script | File read/write | script.md — screenplay in channel character voice |
| Storyboarder | Storyboard | File read/write | storyboard.md — visual sequences, scene layouts, timing |
| Asset Finder | Assets | Web search, file tools | Downloads media resources to episode directory |
| Compositor | Compositing | File read/write, Bash | Episode.tsx — Remotion video composition |
| Producer | Review (all stages) | Read-only | Structured JSON: score (1-10), pass/fail, feedback, upstream issues |
Support Agents
These agents handle tasks outside the main pipeline:
| Agent | Role |
|---|---|
| Content Architect | Episode structure and outline generation |
| Brand Designer | Visual branding and identity |
| Trend Analyst | Trend identification and scoring |
| Channel Wizard | Channel setup and onboarding |
| Curator | Content proposal evaluation |
| Discovery | YouTube trend discovery |
| Idea Router | Idea classification and channel routing |
| Channel Strategist | Long-form channel growth analysis |
How Agents Execute
- The
worker-orchestrateservice determines the next stage to run - It loads the agent definition from
app/agents/{name}.md - Builds a prompt with episode context (channel, episode dir, iteration, prior feedback)
- Invokes the Claude Agent SDK with
--dangerously-skip-permissions(agents run in Docker) - The agent reads the episode directory, does its work, writes artifacts
- The orchestrator checks the result and runs the Producer for review
Producer Review
The Producer agent is the quality gate. After each stage, it:
- Reads the channel's
CHANNEL.mdfor tone and audience expectations - Reads the stage output artifact
- Scores 1-10 with structured JSON output
- Provides actionable feedback if the score is below 7
- Can flag upstream issues (e.g., "the script is weak because the research missed X")
The review uses the Agent SDK's outputFormat: json_schema for guaranteed structured output.
Budget Controls
Each agent invocation has configurable budget limits:
| Stage | Max Turns | Budget | Review Budget |
|---|---|---|---|
| Research | 30 | $2.00 | $0.50 |
| Script | 30 | $3.00 | $0.50 |
| Storyboard | 25 | $2.00 | $0.50 |
| Assets | 30 | $3.00 | — |
| Compositing | 35 | $5.00 | $0.50 |
| Final Review | 15 | $1.00 | — |
TIP
Agent definitions live in app/agents/*.md — each file defines the agent's persona, instructions, and output expectations. Edit them to customize behavior.