# n8n AI agents, and the part nobody demos.

> How I build n8n AI agents that survive production: tool contracts, memory, retries, evals, approval gates, and cost control. Field notes, not a sales page.

Canonical: https://alexcinovoj.com/services/n8n-ai-agents

An n8n AI agent takes an afternoon to demo and a quarter to trust. The gap is tool contracts, memory, evals, and blast-radius limits.

Notes from agents running inside legal, finance, SMB operations, and enterprise IT.

## What an n8n AI agent is made of.

### Tool contracts

Every tool gets a typed input, a narrow permission, and a description the model can act on without guessing. Vague tool descriptions cause more agent failures than weak models do.

### Memory that earns its keep

Short-term state in Redis or Postgres, retrieval over pgvector for the documents that matter. Persist decisions and identifiers, not raw transcripts.

### Bounded loops

Max iterations, max tool calls, and a hard timeout. An unbounded agent is a billing incident waiting for a Friday.

### Approval gates

Human-in-the-loop steps on anything that moves money, contacts a customer, or writes to a system of record.

### Retries and idempotency

Exponential backoff, dead-letter queues on every webhook, idempotency keys on every external write, so a retry never double-charges anyone.

### Traces and evals

Full execution history plus a fixed eval set that runs on every change. If a prompt edit regresses the case set, it does not ship.

## Most n8n agents should have been workflows.

If the steps are the same every time, a deterministic workflow wins on cost, latency, and debuggability. Reach for an agent when the input is unstructured, the path genuinely varies per case, and a human would need judgment to route it. That test kills about half the agent ideas I get asked about, which is the point.

## What actually breaks in production.

Tool descriptions the model misreads. Retrieval that returns plausible but wrong context. Loops that retry a failing tool until the budget is gone. Prompt injection arriving through a document the agent was asked to summarize. Silent success, where the agent reports done and nothing changed downstream. Each of these has a boring, specific fix, and none of them show up in a demo.

## Spend you can attribute to a workflow.

Prompt caching on stable system context, smaller models for classification and routing, the strong model only where reasoning is load-bearing, and a per-workflow cost dashboard. Iteration caps do more for the bill than any model swap.

## Agents that other people operate.

I have shipped n8n agents inside legal, finance, SMB operations, and enterprise IT, then handed them over with a runbook, recorded walkthroughs, and a week of pairing. If nobody can run it without me, it is not finished. Completed 9 Anthropic Academy courses. Lovable Champion and community leader.

## Common questions.

## If the agent works in the demo and nowhere else.

That is a decision problem, not a code-volume problem. The Systems Decision Audit at TechTide AI returns ship, fix, redesign, or stop, with the reasoning written down. When the data underneath the agent lives in Airtable bases, the Airtable Rescue Decision Sprint is the right first move instead.