n8n developer interview questions: what to ask and what a strong answer sounds like

Which screening questions filter out tutorial-only candidates?

Five questions, ten minutes, asked by the hiring manager. Listen for specifics: node names, incidents, fixes.

• Describe the most complex n8n workflow you have run in production. A strong answer names the trigger, the systems involved, the execution volume, what broke and how it was fixed. "I automated a CRM" is a pass.
• Self-hosted or n8n Cloud, and why? Strong candidates know that only self-hosting keeps data inside your network, that it adds patching, backups and upgrades, and that the free Community Edition is licensed under the Sustainable Use License for internal business use, with .ee features behind an Enterprise licence.
• How do you stop a workflow from processing the same event twice? A deduplication key, a store to check it against, and awareness that webhook senders retry on timeouts.
• Where do credentials live in your workflows? In n8n's credential store or an external vault, never in a Code node or expression. Ask who owns the credentials when a contractor leaves.
• What happens when an execution fails? You want to hear "error workflow", "Error Trigger node", "retry on fail" and a notification channel. "I check the executions list" means nobody is watching production.

Which technical questions show real n8n depth?

Eight questions for a 45-minute technical interview with a senior engineer, each with what a strong answer covers.

• When do you use a webhook trigger versus a polling or schedule trigger? Webhooks when the source can push and latency matters; polling or the Schedule Trigger when it cannot or you need batch windows. Strong answers mention test versus production webhook URLs, webhook authentication and the 16 MB default payload limit.
• Explain the item model and how expressions reference earlier nodes. Every node receives and returns an array of items; expressions such as $json, $('Node Name').item.json and $input read from them. Candidates should explain why "run once for all items" versus "run once for each item" changes a Code node's output.
• When do you reach for the Code node? For reshaping arrays, hashing or logic that would take six Set and IF nodes. They should know the Code node runs JavaScript or native Python, that Python runs through task runners since n8n 2.0 with Pyodide deprecated, that npm modules are blocked on Cloud, and that HTTP calls belong in the HTTP Request node.
• How do you split a large workflow into sub-workflows? With the Execute Sub-workflow node and the "When Executed by Another Workflow" trigger with a defined input schema, for reuse, memory limits, independent testing and clearer ownership.
• Walk me through your error-handling layers. Node-level Retry On Fail, the On Error setting (stop, continue, or continue using the error output), a Stop And Error node for business-rule failures, and a workflow-level error workflow that receives the execution ID, failed node and stack trace. Bonus points for separating transient errors (retry) from permanent ones (alert and park).
• How do you handle rate limits and pagination? Batching with Loop Over Items or Wait nodes, backoff on 429 responses, cursor or offset pagination with a termination condition, and never loading every page into one execution.
• Design an idempotent writer for a CRM. Look up by external ID before creating, use upserts where the API supports them, store processed event IDs, and run the dedupe check before any side effect.
• How does the AI Agent node work, and when would you not use it? The AI Agent node is built on LangChain, takes a chat model, optional memory and at least one tool sub-node, and can connect to MCP servers as tools. Strong answers limit tools and permissions, cap iterations, log tool calls, and prefer a deterministic workflow when the task needs no reasoning.

What practical exercise should you set?

Give the candidate a sandbox instance and 60 minutes. Scenario: an e-commerce platform posts order events to a webhook and retries any request that gets no 200 within ten seconds. For each order, look up the customer in a CRM, create the contact if missing, post a summary to Slack, and append a row to a database. The CRM returns 429 under load and the database is sometimes down.

What to look for: the webhook responds immediately and hands work to a sub-workflow rather than making the sender wait; a dedupe key on the order ID stops retries creating duplicates; the CRM step retries with backoff; the database step continues using the error output into a parking table or alert; and credentials come from the credential store. A happy path with no error handling fails the exercise, however tidy the canvas.

Which production and evaluation questions matter?

Five questions for senior and lead candidates, or anyone who will own your instance.

• How do you scale n8n beyond a single process? With queue mode: the main instance handles triggers and webhooks, Redis queues execution IDs, workers run executions with a default concurrency of 10, and webhook processors can be added behind a load balancer.
• How do you monitor executions? Error workflows for alerts, the executions list for debugging, Insights for failure rate and run time on paid plans, and log streaming on Enterprise.
• How do you version workflows and promote between environments? n8n's Git-based source control connects an instance to a repository with one branch per environment; n8n's pricing page lists it on Business and Enterprise plans as of October 2026. On lower tiers, strong candidates export workflow JSON to Git and use variables for environment-specific values.
• How do you keep personal data out of execution logs? Execution data stores every item that passed through a node, so a GDPR-aware answer saves on error only, disables progress data, sets pruning with EXECUTIONS_DATA_MAX_AGE (default 336 hours), hashes or drops fields before logging, and keeps a map of which workflows touch customer data.
• How do you test an AI workflow before it goes live? Evaluation sets with expected outputs, n8n's evaluations feature, pinned data for repeatable runs, cost caps per execution, and a human review step before anything writes to a customer-facing system.

What are the red flags?

• API keys pasted into Code nodes, expressions or workflow notes.
• No error workflow on anything they have shipped, or alerts sent to a channel nobody reads.
• No deduplication key and no answer when you ask what a retried webhook does.
• Plans to resell a hosted n8n to clients without having read the Sustainable Use License.
• Treats the AI Agent node as magic: unlimited tools, no iteration cap, no logging of tool calls.

How do you run the interview loop?

Three rounds in five working days. Day one: a 20-minute screen by the hiring manager. Day two or three: the 60-minute build exercise, observed by an engineer who has shipped n8n. Day four: a 45-minute technical conversation that starts from the exercise and adds the production questions for senior candidates. Decide within 24 hours.

Score every round on the same four axes: correctness, production habits (idempotency, errors, secrets), communication, and judgement about when not to automate. The vetting guide covers scorecards and panel design in more depth.

Or skip the loop

Without an engineer to observe the build exercise, this loop is hard to run well. Hyparz matches pre-vetted remote n8n developers typically within days; about 5% of AI-role applicants pass vetting, which covers the production habits above. Engagements start with a one-week risk-free trial and you keep any work delivered. See hire n8n experts, AI automation services, or the companion n8n developer job description template if you are hiring in-house.