Multi-Agent Lead Workflow: Separating LLM Extraction from Deterministic Routing
aiminent_ai · reddit · 2026-08-10
The author shares a multi-agent lead qualification workflow built for the real estate industry, covering extraction, enrichment, scoring, property matching, and follow-up.
Core Architecture Design
- Process Breakdown: Incoming raw enquiries pass sequentially through extraction, enrichment, scoring, conditional routing, and matching/follow-up.
- Responsibility Division: The author highlights a key design decision—limiting the LLM's role to understanding, extracting, and classifying (outputting structured JSON), while delegating actual scoring calculations and business routing to deterministic workflow logic.
This separation makes the system more predictable and allows scoring weights to be adjusted without modifying prompts. The author asks the community: when building agents, where exactly should we draw the line between model reasoning and deterministic business logic?
More from coding & agent
- How Do Developers Decide Agent Permissions in Prod? Mostly Vibes — sp_archer_007 · 2026-08-10
- 11k-Star Open-Source Intel Terminal Crucix Integrates LLMs for Auto-Analysis — tom_doerr · 2026-08-10
- i-have-adhd Plugin Forces Coding Agents to Output Action First, Zero Fluff — aigclink · 2026-08-10
- Secure Credential Storage for AI Agents in Production Environments — ashiflyy · 2026-08-10
- Firecrawl Open-Sources Web Data Agent for Structured Web Research — tom_doerr · 2026-08-10
- Open Source PaaS Coolify Hits 40k Stars, Offers Free Self-Hosted Vercel Alternative — thisdudelikesAI · 2026-08-10