a builder's codex
codex · operators · Hamza Farooq · ins_three-class-agent-taxonomy

Agents come in three classes — tag each loop or under-resource it

By Hamza Farooq · AI engineering practitioner; co-author with Jaya Rajwani in Lenny's Newsletter · 2026-04-28 · essay · Not All AI Agents Are Created Equal

Tier B · TL;DR
Agents come in three classes — tag each loop or under-resource it

Claim

"Agent" describes three structurally different systems with order-of-magnitude different build costs — deterministic automation (n8n, Zapier), reasoning-and-acting agents (LangGraph, CrewAI), and multi-agent networks — and treating "agent" as one estimate causes systemic mis-resourcing on roadmaps.

Mechanism

Deterministic automation has a fixed graph and no model reasoning step; build cost is hours to days. Reasoning-and-acting agents have a model-driven tool loop; build cost is weeks. Multi-agent networks coordinate multiple agents with emergent behavior; build cost is months because coordination, observability, and failure modes compound. A roadmap that treats all three as "build an agent" estimates the deterministic cost and ships nothing on the multi-agent side. Tagging each agent loop with its class lets resourcing match the cost band.

Conditions

Holds when:

Fails when:

Evidence

"One agent might take six weeks to build. Another might take six months."

Class breakdown:

1. Deterministic automation (n8n, Zapier) — fixed graph, no reasoning step.

2. Reasoning-and-acting agents (LangGraph, CrewAI) — single agent with a tool loop.

3. Multi-agent networks — multiple agents coordinating, with emergent behavior.

— Hamza Farooq and Jaya Rajwani, https://www.lennysnewsletter.com/p/not-all-ai-agents-are-created-equal, 2026-04-28

Signals

Counter-evidence

The taxonomy is taxonomic and edge cases blur. Some operators argue the three classes overlap so heavily that the distinction is academic; what matters is observable cost per agent over time. The taxonomy is most useful for greenfield planning, less for live-system tuning.

Cross-references

Open the interactive view → View original source → Markdown source →