Claude Code Routines (and Managed Agents) vs n8n: Deterministic or Agentic Automation?

Anthropic launched Claude Managed Agents on April 9, 2026, then Claude Code Routines on April 14. Both run a Claude agent in the cloud, on a schedule or on an event, with no server to manage. Since then, our clients keep asking the same question: does this replace n8n?

No, and it is not the same kind of tool. n8n runs a sequence of steps you drew. A Routine runs a natural-language instruction and decides the steps itself. The first approach is deterministic, the second is agentic.

This article compares them on what matters in production: control, cost, limits, security. With a real n8n workflow we built for Legalstart, and the exact spot where an agent would earn its place.

Facts checked on September 7, 2026. Routines are in research preview and Managed Agents in public beta: quotas and prices can change.

The short answer

Keep n8n as the backbone for everything that must run identically every time: receiving documents, extraction, syncing tools, writing to a database. You see every step, you replay a failed run, you know what it costs.

Use a Claude Code Routine or a Managed Agent for the steps that require reading, understanding and deciding: triaging alerts, reviewing a pull request, writing a weekly summary from scattered sources, handling the ambiguous case that breaks the workflow.

Then plug the second into the first. n8n triggers the agent over its API when a step needs judgment, the agent returns a result, n8n takes over again. That is the architecture we recommend in 2026.

Deterministic or agentic: the difference that matters

An n8n workflow is a graph. A trigger, nodes, connections. Every run follows the same path with the same conditions. If step 4 fails, you see it at step 4, with the exact input that failed. Even when a node calls an LLM, you decided when the model steps in and what it is allowed to touch.

A Claude Code Routine is a prompt, one or more GitHub repositories, MCP connectors and a trigger. On every run, Claude reads the prompt and decides what comes next: which files to open, which commands to run, which tool to call, when to stop. The output can vary from one run to the next, because the agent adapts to what it finds.

So the question is not « which one is more powerful ». It is: who should decide the steps, you or the model? For high-volume processing with known rules, you. For a task whose rules cannot be written in advance, the model.

n8n in 2026: what you are actually buying

n8n is no longer the small open-source tool of 2021. A $180M Series C in October 2025 led by Accel at a $2.5B valuation, more than 200,000 GitHub stars, and a stated ambition to be « the platform to build with AI ».

On pricing, n8n Cloud starts at €20 per month on annual billing for 2,500 executions (Starter), €50 for 10,000 (Pro) and €667 for 40,000 (Business, with SSO, environments and a self-hosted option). An execution is one full run of the workflow, whatever the number of steps: that is what keeps the cost predictable. Self-hosted, the Community Edition is free under the Sustainable Use License: internal use and client work are allowed, reselling n8n as a hosted service is not. It is not open source in the strict sense, and you should know that before installing it at a client.

On AI, n8n ships an AI Agent node with tools and memory, MCP Server Trigger and MCP Client Tool nodes, Evaluations to test model outputs, and Data Tables. In other words, n8n can do agentic work, but inside a frame you draw. We covered what MCP changes when connecting AI to business tools in a dedicated article.

What you have to accept: an n8n workflow needs maintenance. Every use case is one more node, every third-party API change can break a connector, and self-hosting means following near-weekly updates.

Claude Code Routines and Managed Agents: what they actually are

Routines are Claude Code tasks configured once (prompt, repositories, connectors) and executed on Anthropic’s infrastructure. Three triggers: a schedule (hourly at most), an API endpoint you call with a POST and a token, or GitHub events (pull request, release). They run without human approval: no permission mode, no question asked along the way. By default, commits go to branches prefixed claude/ under your own GitHub identity.

Quotas depend on the subscription: 5 runs per day on Pro, 15 on Max, 25 on Team and Enterprise, with extra usage billed beyond that. A Routine draws on your Claude usage like an interactive session. Every run starts from scratch with no state carried over, and during the research preview a Routine belongs to a personal account, with no team sharing.

The same mechanism exists for non-developers in Claude Cowork, as scheduled tasks: daily briefing, weekly report, competitor monitoring.

Managed Agents is the layer above, for teams building their own agent: a hosted runtime with sandboxing, long-running sessions, multi-agent coordination and tracing. Billing combines token prices and $0.08 per active session-hour. Notion and Asana are among the first customers Anthropic cites.

What you have to accept: you do not look at a graph, you read a trace. Cost depends on the reasoning consumed, not on the number of runs. And the agent acts with the rights you gave it, on inputs you do not always control.

Comparison table: n8n vs Routines vs Managed Agents

Real case: Legalstart’s Kbis workflow, and where an agent would fit

For Legalstart, we built an n8n workflow that picks up Kbis extracts (the French company registration certificate) received as PDFs in a mailbox, sends them to an OCR service, extracts the SIREN number, the company name and the other useful fields, then stores the PDF on S3 and the data in a database.

Why n8n and not an agent: every step is known in advance and must run identically on every document. A Kbis misread by the OCR is an error at step 3, visible, with the file that caused it, and an error workflow that sets the document aside. The cost is the same for 50 or 5,000 documents a month. And nobody wants an agent improvising on how to write a SIREN into a database.

Where an agent would fit: the documents set aside. A skewed scan, a multi-page Kbis, an ambiguous field. Today, a human picks those up. A Routine triggered by n8n through its API endpoint, with the PDF and the partial OCR result as context, could propose a correction, explain it, and send it back to the workflow for validation. The deterministic part handles the vast majority of documents, the agentic part handles the rest, and a human keeps the last word on what goes into the database.

That is the general pattern: the agent is not the workflow, it is one step of the workflow. We had already seen it while building an ops agent that keeps tickets, Slack, Notion and GitLab in sync: the real work is orchestration and side effects, not the model.

When to choose what

The traps on both sides

On the Routines and agents side, the number one topic is security. A Routine runs without approval, with your Slack, GitHub or Drive connectors, and reads inputs you do not control: a pull request description, the text sent to the API endpoint. A prompt injection in an external PR can steer the agent. Minimum rules: a dedicated account rather than your personal identity, only the connectors you strictly need, and never write access to production without downstream validation.

Second trap: quotas. 5 or 15 runs a day get used up fast once GitHub events are wired in, and events beyond the cap are dropped, not queued. With Managed Agents, a session stuck in a loop burns tokens and hours.

On the n8n side, the classic trap is the workflow that grows until nobody can read it, with business logic hidden in Code nodes. Then the neglected self-hosted instance: no backups, no updates, hard-coded credentials. And the cloud execution count: a trigger firing every minute burns through more than 43,000 executions a month before doing anything useful.

What we recommend at Fenxi

For an SME or mid-sized company automating in 2026:

• n8n as the backbone, self-hosted if someone can maintain it, cloud otherwise;
• AI steps inside n8n first (AI Agent node or an API call to Claude) as long as they stay bounded: classify, extract, summarize;
• a Claude Code Routine or a Cowork scheduled task for recurring work that needs judgment and produces a human-readable deliverable;
• Managed Agents only if the agent is part of your product;
• a dedicated technical account for agents, minimal rights, human validation before any irreversible write;
• one execution log for everything, agent or not, so you know what ran and what it cost.

Conclusion

The « n8n or agents » debate is the wrong question. The right one: where in your process do you need judgment, and where do you need the exact same thing to happen a thousand times? Draw the second in n8n. Hand the first to an agent, plugged into n8n, with limited rights. You get the predictability of a workflow and intelligence where it is useful.

Automating with n8n and wondering where AI belongs?

Describe your process. We can audit your workflows, identify the steps that justify an agent, and build the architecture that holds up in production.