n8n's AI Agent can lie to you — and still show green

The exact clicks

1. Quick term check: in n8n, a "node" is one block in your workflow — you drag it in from the panel on the left and wire it to the blocks before and after it. Every "node" below is one of these blocks.
2. This fix has two parts — a workflow-level error trigger alone does not catch this, because the AI Agent resolves the failure internally before the workflow ever "errors." Both parts are needed.
3. Part 1 — make every tool report its own status honestly. Open each Tool node your AI Agent calls. For an HTTP Request tool, go to Options → Response and enable "Never Error" — this stops the tool from throwing silently and lets you inspect what actually happened instead.
4. For a Code node or sub-workflow used as a tool, wrap its logic so it always returns one structured object instead of throwing: {"status": "ok" | "failed" | "empty", "data": ...} — a plain success/failure field the next step can actually read.
5. Part 2 — add a real check after the Agent. Add an IF node (or a Code node) directly after your AI Agent node.
6. In that node, check the Agent's own output for the failure signs: any tool result where status equals "failed", a tool missing entirely from the Agent's intermediate steps, or a result with zero items where you expected real data.
7. Route the "broken" branch of that IF node into a Stop And Error node — this is what actually makes the workflow fail for real, instead of silently continuing on bad data.
8. Only now does the standard fix apply: create a small separate workflow starting with an Error Trigger node (no configuration needed) — this is where you add a Slack/email/Discord alert.
9. In your main workflow (the one with the Agent), open the three-dot menu → Settings → Error Workflow, and select the error-handler workflow you just made. Save.
10. Test it honestly: temporarily break one tool on purpose (a wrong API key works well) and run the workflow for real. If you did this right, you get an actual alert — not a quiet green checkmark.

What changes

• A real, sourced defect: n8n's own GitHub has two separate open issues on this (#22771, #24042), and an active 2026 community thread from someone already on n8n's latest version — this isn't a stale, already-fixed bug.
• n8n treats a tool's failure as part of the AI Agent's own reasoning process, not as a workflow execution error — so the built-in Error Workflow mechanism never fires on its own, no matter how it's configured.
• The real fix isn't the Error Trigger alone — it's making every tool report status honestly, checking that status right after the Agent, and deliberately throwing before anything destructive happens.

What it will not do

• This is a design characteristic of n8n's current AI Agent node, confirmed as unresolved as of this episode's publishing (one fix proposal was closed "not planned" in March 2026) — re-check n8n's own GitHub if this episode is more than a few months old.
• The fix adds real setup work per tool, per agent — it does not come free, and skipping it on even one tool leaves that one silent.
• This episode covers one specific, documented failure mode (a tool call failing silently inside an Agent) — not a general audit of n8n's reliability.