MULTI-AGENT FABLE 5 ORCHESTRATION

ARCHITECTURE OVERVIEW
Main orchestrator (Kai) routes tasks to 3 specialized sub-agents
→ Agent A: Research and analysis
→ Agent B: Writing and content creation
→ Agent C: Code generation and debugging
Memory layer: Hermes (vector DB for context retention)
Model routing: Claude Opus for orchestrator, Claude Sonnet for sub-agents (cost optimization)

OPERATION FLOW

1. USER INPUT → ORCHESTRATOR
   Kai receives task: "Write a blog post about Python best practices"
   Kai analyzes intent and extracts metadata
   Decision: route to Agent B (writing) + Agent A (research) in parallel

2. ORCHESTRATOR → SUB-AGENTS
   Agent A spawned with instructions: "Research Python best practices, focus on async/await"
   Agent B spawned with instructions: "Write structured blog post outline from research"
   Both agents write outputs to shared memory (Hermes vector DB)

3. SUB-AGENTS COMPLETE
   Agent A output: research notes + 5 key sources
   Agent B output: blog post draft + SEO keywords
   Both outputs stored with embedding for future reference

4. ORCHESTRATOR REVIEWS
   Kai reads both outputs
   Checks quality gates (post > 2000 words, sources are recent)
   Routes to Agent C if code examples need verification
   Agent C: "Fix and test any Python code snippets"

5. FINAL OUTPUT
   Kai assembles final blog post from Agent B + verified code from Agent C + sources from Agent A
   Returns to user
   Logs conversation to memory for future context

MEMORY PERSISTENCE (HERMES)
Stored as vector embeddings with metadata tags:
• previous research on Python topics
• user's preferred writing style
• common technical errors (code patterns to avoid)
• successful blog structures

When new task comes in, Hermes retrieves similar past work so agents don't reinvent wheels.

MODEL SELECTION AND COST OPTIMIZATION
Main orchestrator (Kai): Claude Opus (fast reasoning, complex routing)
Research agent: Claude Sonnet (balanced cost/speed)
Writing agent: Claude Sonnet (creative tasks, lower cost than Opus)
Code agent: Claude Opus (safety-critical, debugging complexity)

Estimated monthly cost: ~200 multi-agent tasks = 1200 (vs 2400 if all used Opus)

AGENT CONFIGURATION (SKILL.MD EXCERPT)
Specify for each agent:
→ Role (researcher, writer, coder)
→ Guardrails (what they can't do)
→ Output format (structured JSON for routing)
→ Error handling (when to escalate to orchestrator)
→ Memory access (read-only vs read-write for Hermes)

DEBUGGING AND MONITORING
Log each agent's decision and output
Track latency (research should complete in 30 sec, writing in 60 sec)
Alert if any agent fails 2x in a row
Measure quality (user feedback on final output)