Multi-Agent Job Outreach Pipeline — Supervisor/Worker Pattern in Python + Claude

Wanted to share a multi-agent pattern I've been running — it's a supervisor/worker architecture that maps cleanly to n8n if anyone wants to port it.

The Pipeline

Agent 1 — Ingestion + Scoring

Pulls from Remotive API, WeWorkRemotely RSS, and Upwork RSS (all free, no keys needed)
Each item goes to Claude with a structured scoring prompt (1–10 against defined criteria)
Only items scoring ≥ 6 pass through to the next agent
Output: items_YYYY-MM-DD.json with score, reason, pain point, and pitch angle per item

Agent 2 — Context-aware Email Drafter

Reads the scored items JSON
For each item: Claude reads the full source text and writes a personalised email that references specific details — stack, stated problems, their own wording
No templates. Every draft is generated fresh against the actual source content
Output: drafts_YYYY-MM-DD.json — human reviews before anything is sent

Agent 3 — Gmail MCP Send

Approved drafts sent via Gmail MCP directly from inbox
Full send log maintained in JSON

Why This Pattern Is Interesting Technically

The scoring step is the key. Without it, you're passing 60+ low-signal items to an LLM and burning tokens on noise. With a Claude scoring gate, only ~10–15% of raw items make it through — and those are the ones worth spending a full generation on.

The other thing worth noting: keeping state in flat JSON files rather than a DB is underrated for single-user pipelines at this scale. Simple to inspect, easy to debug, no infra overhead.

n8n Equivalent of Each Agent

Agent	n8n Implementation
Agent 1	Schedule trigger + HTTP Request nodes + AI Agent node (scoring prompt) + Filter node
Agent 2	Manual trigger + Loop over items + AI Agent node (drafting prompt) + Write file
Agent 3	Manual trigger + Gmail node with approval gate