AI Agents in n8n: Where No-Code Breaks

Three signals converged on the n8n forum almost simultaneously.

A bug in HTTP Request node version 2.38.5: the node receives valid JSON from another workflow's webhook, but substitutes template strings such as `{{ $json.property }}` instead of real values. Subsequent nodes receive text with curly braces instead of data.

Nearby is a fresh freelancer listing: a fixed £95 to diagnose one n8n workflow of up to 30 nodes, with static checks for broken connections, disabled nodes, unauthorized webhooks, and hardcoded secrets. The n8n blog also features articles on the reflection pattern and why agents cannot be built through prompting alone.

All three signals point to the same gap: an agent is easy to assemble, but far harder to make reliable enough for production.

No-code builders such as n8n win at the first step: assembling an agent that calls a couple of APIs and a model takes hours.

Time to the first result is minimal, which is exactly why such tools quickly reach sales or support teams. The problem starts after that.

The re-templating bug stays silent: the workflow does not fail or report an error; it simply passes garbage instead of data, revealed when a client receives an email containing `{{ $json.name }}` instead of their name.

The paid-audit listing identifies the same classes of defects: no retries for external calls, a webhook accepting requests without signature validation, and a secret stored in plain text in the workflow export.

None of these defects is visible during a demo run.

In production, each of them can lead to a data leak, a lost lead, or a botched email to a client.

The n8n blog puts it plainly: the reflection pattern—an agent generates, critiques, and revises its answer—is one of four core agentic patterns, alongside tool use, planning, and multi-agent scenarios.

The idea works, but it has a cost that another article on the same blog states directly: the more often you ask a model to evaluate its own output, the more opportunities it has to hallucinate again.

Having an LLM check itself with the same LLM requires additional model calls and creates another chance to make a mistake in the same place where it already failed.

This is the view of n8n practitioners, not supported by a separate study, but it matches what live workflows show: an agent that praises itself can go through two rounds of reflection and still output the same curly braces instead of a value because the error was in parsing by a node that ran before the model generated anything.