Replacing n8n with Claude Code and a CLAUDE.md file
Workflow builders are great until your workflow is "think for four minutes, then write 2,000 words."
Our long-form content pipeline kept dying on API timeouts. So I threw out the workflow builder and gave an agent persistent memory instead. Here's the architecture, the failure modes, and the numbers after 30 days.
The core problem with node-based execution environments is their rigidity in the face of non-deterministic processes. When an LLM decides it needs to execute an external tool, fail, re-read documentation, and try again, modeling that in a DAG becomes an exercise in drawing endless spaghetti webs of error-handling nodes.
Instead of telling the system how to move data, we shifted to telling it where the rules are stored. The entire orchestration layer collapsed into a single shell command and a perfectly tuned markdown file.
Where the pipeline kept breaking#
Initially, we had a 40-node setup. Extract RSS, clean text, summarize, draft sections, compile. Every API hiccup meant manual intervention. We needed a way to provide context to a persistent agent session. The solution was surprisingly brutalist.
# The new orchestrator is literally just this
cat payload.json | claude "Execute the routine defined in CLAUDE.md for this data. Output only the final markdown file path."
The magic isn't in the command; it's in the memory structure. By keeping a local file that the agent is forced to read on initiation, we offload process state from the runner to the agent itself.
"An agent with a memory file beats a workflow with twenty nodes, but only if you version the memory."
To visualize the difference in complexity, consider the architecture delta:
We cut infrastructure costs by 80% and execution reliability jumped to 99.4%. The only downside is that you have to write system prompts that are basically legal contracts.