We run AI workflow automation in production every week, and we publish what we learn here. AI Workflow Lab is a hands-on journal for engineers wiring up n8n, LangGraph, and Model Context Protocol agents into systems that actually ship — with the timeout retries, observability, and cost ceilings that separate a demo from a Tuesday-morning automation handling 40,000 invoices.
n8n stopped being just a Zapier alternative the moment it shipped native AI Agent and Tool nodes. Our team builds workflows where a single agent node reasons over a webhook payload, calls three sub-workflows as tools, and writes structured JSON back to Postgres — all without a Python service in the loop. The traps are real: token budgets blow up when you forget to summarise long memory, queue mode behaves differently from main mode under load, and the official queue mode docs are required reading before you push past a few hundred executions an hour.
We cover the boring infrastructure most tutorials skip: pinning a self-hosted instance to a specific n8n version, backing up the SQLite or Postgres state, wiring OpenTelemetry traces out of custom nodes, and the exact pattern we use for human-in-the-loop approval gates that survive a worker restart.
LangChain's chain-of-runnables abstraction was never going to scale to multi-step agents with branching, retries, and shared state. LangGraph is what we reach for now — explicit nodes, explicit edges, and a checkpointer you can actually inspect. We write up the patterns we use in shipped products: routing supervisors, parallel tool fan-out with a reducer, durable execution against Postgres checkpoints, and how to keep a graph runnable from both a FastAPI endpoint and a background Celery worker without rewriting it twice.
Where the docs go quiet, we go loud. Streaming partial state to a Next.js client over Server-Sent Events, replaying a checkpoint after a model swap from Claude 4.7 to GPT-5, and the evaluation harness we plug into OpenAI Evals so a graph change can't silently regress accuracy on the regression suite.
The Model Context Protocol is the piece most teams underestimate. Once your Claude desktop, Cursor, and internal agent runtime all speak MCP, you stop writing one-off tool wrappers and start treating company systems — Jira, Snowflake, the internal billing API — as servers that any compliant client can discover. We document how to stand up an MCP server in TypeScript or Python, scope it with OAuth, and keep the tool schemas tight enough that an agent doesn't hallucinate a parameter that costs you money.
We also publish the unglamorous side: writing a Dockerfile that runs an MCP server behind a reverse proxy, rate-limiting noisy clients, and the audit trail we keep so security can answer "which agent touched this row?" six months later.
The articles below are the working notes — failure modes, copy-pasteable configs, benchmark numbers from our own staging cluster. Browse the latest below, or jump straight to the tag that matches whatever's on fire this week.
A practical, evaluation-first guide to shipping text-to-SQL with LLMs in production: schema linking as RAG, semantic layers, three-stage query validation, and defense-in-depth guardrails, with runnable code and 2026 benchmark numbers.
Voyage-3-large tops MTEB v2, Cohere Embed v4 handles 128K multimodal, OpenAI is the best default, and Jina wins multilingual. Full 2026 benchmark with code.
OpenAI Batch API and Anthropic Message Batches both cut LLM token costs by 50% in exchange for a 24-hour SLA. Here is when batch wins, working Python for both, and the production pitfalls nobody mentions.
When to pick structured outputs vs function calling across OpenAI, Anthropic, and Gemini in 2026: schema enforcement, latency, and runnable code for each pattern.
Compare Inngest, Trigger.dev, and Hatchet for AI workflows in 2026. Durable execution, retries, pricing, self-hosting, and which queue fits your stack.
An honest 2026 comparison of LiteLLM, OpenRouter, and Portkey across routing, caching, observability, governance, and pricing, with real configs and a decision framework for picking the right gateway.
A field-tested walkthrough for wiring an n8n instance into Claude Code over MCP, covering server install, tool registration, environment variables, and the debugging steps that actually unblock the stdio handshake.
A field-tested walkthrough of why n8n AI Agent nodes recursively call themselves when tool descriptions overlap, plus three prevention patterns I now apply to every production workflow.
I lost most of a Tuesday to an n8n AI Agent node that flat-out refused to call the HTTP Request tool I'd wired into it. After pulling apart a dozen variations I have six reproducible causes.
How to chunk documents for RAG in 2026: contextual retrieval, late chunking, semantic splitting, plus production Python code, benchmarks, and evaluation tips.
Compare Mem0, Letta, and Zep on architecture, latency, recall accuracy, and cost. Includes LangGraph integration patterns and production pitfalls to avoid.
Pure vector search fails on roughly 40% of real-world queries. This 2026 production guide shows how to combine BM25 sparse retrieval, dense embeddings, RRF fusion, and a cross-encoder reranker into a single pipeline that pushes recall@10 from 78% to 91% — with working Python code.
Choose your preferred language to explore our content