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:
- The team is shipping more than one agent and needs prioritization.
- Build estimates have meaningful business consequences (deadlines, headcount allocation).
- Engineering can distinguish the three classes (or willing to learn).
Fails when:
- The team is shipping one prototype and the taxonomy is overhead.
- The category genuinely fits one class only (e.g., pure deterministic ops automation).
- Class boundaries blur in practice — frameworks like LangGraph allow both reasoning and multi-agent shapes.
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
- Roadmap items have an explicit agent-class tag, and resourcing matches.
- Build estimates separate the three classes rather than averaging.
- Multi-agent networks get longer observability and failure-mode investment than single-agent loops.
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
- Agents work when treated as a team, not a single super-tool — Claire Vo's operating model lives in the multi-agent class but works through identity discipline.
- Map agents 1:1 to enumerated jobs-to-be-done, not abstractly to "AI-augmented" workflows — adjacent: the spec layer that should precede class selection.