Job Agent — Multi-Agent Architecture Notes

Architecture-focused notes for the [[../career-strategy/job-application-agent|Job Application Agent project]]. The other note has the project overview; this one is the technical pattern reference.

Pattern: 5-Robot Multi-Agent System

Layer 1: Producers (Scrapers)

Five independent scrapers run on schedule (4-6h) and produce job postings. Decoupled — adding a new scraper doesn't change the rest of the pipeline.

Layer 2: Filter (Fit Scorer)

Single AI gatekeeper. Receives all jobs, applies hard rules (salary floor) + soft scoring (skills/role/growth). Below threshold → drop. Above → forward.

Key insight: Filter BEFORE expensive operations (CV tailoring, browser automation). Don't waste 60s of Qwen inference on a job that's clearly below salary floor.

Layer 3: Generators (CV Builder)

Two AI calls per accepted job:

1. CV tailor (~30s, 2000 tokens output)
2. Cover letter (~15s, 800 tokens output)

Files persisted to disk so executors can upload them.

Layer 4: Routers (Platform-specific Executors)

Three executor agents — one per platform tier:

• India Agent — auto-submit OK
• Abroad Agent — auto-submit OK, always include cover letter
• Tier 1 Agent — fill but DO NOT submit (high-stakes companies, manual review)

The router pattern lets us add new platforms without touching existing executors.

Layer 5: Sink (App Logger)

Single endpoint that writes structured rows to Google Sheets. Decoupled — every other component just POSTs to /log-application regardless of what happened.

Key Architectural Decisions

Why local Ollama instead of cloud API

• Cost: ₹0 vs ~₹500/month at Anthropic prices for ~50 jobs/day
• Privacy: CV data never leaves the laptop
• Speed: ~15s per scoring call is acceptable for an async pipeline
• Quality: Qwen 2.5 7B reliably outputs structured JSON with format: "json" mode

Why Playwright instead of Selenium

• Built-in waits (no flaky time.sleep())
• Modern auth state persistence (cookies survive between runs)
• Better screenshot API (full-page support)
• Bundles Chromium (no driver version mismatch)

Why webhooks instead of direct function calls

• Each robot is independent (can be debugged/disabled in isolation)
• n8n's execution log preserves the full trace per job
• Future: can move any robot to a different machine without touching others

Why JSON mode (not free text)

• Qwen with format: "json" returns valid JSON ~99% of the time
• Saves a parsing-error retry loop
• Trade-off: slightly higher token count, marginally slower

Bottlenecks Observed

| Bottleneck | Mitigation | |---|---| | Ollama cold start ~30s | Keep-alive ping every 4 minutes | | LinkedIn anti-bot detection | Saved auth cookies in linkedin_auth.json | | Tier 1 ATS form variation | Comprehensive selector list (10+ aliases per field) | | n8n single point of failure | Soul in Motion auto-starts via scheduled task |

Patterns to Apply Elsewhere

This same 5-layer pattern can power:

• Content automation — scrape ideas → score → write → publish → log
• Lead qualification — scrape leads → score → enrich → outreach → log
• Investment screener — scrape filings → score → research → alert → log
• News digest — scrape sources → score → summarize → email → log

The shape stays. The prompts and executors change.

Related

• [[../career-strategy/job-application-agent]] — project overview
• [[_index]] — broader AI agent learning notes