May 12, 2026 · 14 min read

How we built Lonesyte's 12-agent orchestrator: Core MVP in 4 weeks

Twelve specialist agents over a custom orchestration layer that replaced 16 hrs/wk of SDR triage. Why we picked LangGraph, where the Assistants API would have broken under load, and the cost-routing trick that kept the bill 7× cheaper than naive GPT-4.

AI AgentsProduction

A network of connected nodes representing a multi-agent orchestration graph

Most multi-agent demos break the moment they meet real traffic. Lonesyte's didn't. The reason was a boring orchestration layer that does three things a production agent system needs: bounded retries, structured handoffs, and model routing.

The problem brief

Najeeb's SDR team was burning ~40% of their week on triaging inbound. Three CRMs, two enrichment vendors, none of them talking to each other. The founders had a Notion page that said 'AI' and a six-month runway. We had four weeks to ship the Core MVP (Phase 1) they could put in front of paying customers. They've kept iterating since, the product is on v5 now.

Why LangGraph and not the Assistants API

Three reasons. First, the Assistants API hides state in OpenAI's infra, when (not if) a tenant needs to audit what an agent saw before it acted, you're stuck. Second, tool calls in the Assistants API serialize through OpenAI's queue, which becomes the bottleneck at ~30 concurrent. Third, model routing requires owning the call graph: when 80% of your reasoning is mid-grade and only the final synthesis needs frontier, you save ~85% on inference if you can swap mid-graph.

The 12 agents (and what they actually do)

→Qualifier, yes/no lead-fit pass against a learned spec
→Enricher, HubSpot + Apollo + warehouse lookup, dedup'd
→Router, assigns owner by territory + workload
→Drafter, first-draft outreach, founder-tuned tone
→Critic, second-pass on tone + factual claims (catches 60% of bad drafts)
→Logger, writes back to CRM with structured deltas
→...and 6 more specialists for edge cases and recovery

Cost routing in practice

The qualifier and logger run on Haiku-class models, sub-cent calls. The critic runs on Sonnet-class. Only the drafter ever touches frontier, and only when the critic flags low confidence. Average cost per triage dropped from $0.42 with GPT-4 everywhere to $0.06 with routing. At ~40k triages/month that's the difference between a $17k inference bill and a $2.4k one.

What broke (and what we kept anyway)

The eval harness took longer than the agent code. We shipped one, and few teams discuss this part of building a production agent system. Happy to share the template if you ask.