Build a AI agent workflow with n8n + Claude API + webhooks
Autonomous agents that handle multi-step tasks — triage, drafting, classification, scheduled work.
Agents that actually run in production
"AI agent" usually means a fragile loop that breaks the second the LLM's output is malformed JSON. Production agents need validation, retry logic, escalation paths, and observability. My full write-up on the architecture.
What I build
- n8n orchestration — the workflow is visual, every step has an inspectable input/output, your team can debug without me.
- Strict JSON output prompts — Claude returns structured data with a discriminated union for the decision (send / skip / escalate). Validated server-side; bad output is retried with a recovery prompt.
- Escalate path — when the agent is uncertain, it routes to a human queue instead of guessing.
- Cost guardrails — per-workflow daily token budget, Slack alerts at 80%, hard stop at 100%.
- Loop prevention — every output is stamped with the agent name; agents refuse to process inputs they generated themselves.
FAQ
How is this different from Zapier with AI steps?+
Zapier's AI steps are a thin LLM call inside their workflow runtime. The pattern I use treats the LLM as a structured-output decision engine with explicit guardrails. The result is dramatically lower hallucination rates and clearer failure modes.
Can the agent take real actions (send emails, update databases)?+
Yes, but with explicit tool definitions and idempotency keys. Every external write is logged and reversible.
What use cases work best?+
Triage (sorting inbound), drafting (first-pass content), classification (tagging records), monitoring (watching feeds and pinging on signals). Agents are bad at multi-hour open-ended planning — that's what humans are still for.
Ready to build a AI agent workflow?
Tell me what you're shipping. Reply within 2 hours during EU business hours.