n8n AgentsAgents are now first-class in n8n

Agents are first-class artifacts that live alongside your workflows in a project. An agent has a model, instructions that define its role and constraints, tools it can call, triggers that start it, sessions that record its conversations, and configurable memory. You build one by describing it to the AI Assistant or by configuring it manually in the Agent Builder, then you test the draft, publish it, and run it. At each step the model decides what to do next — reply, look something up, escalate, or finish — within the tools you gave it.

The AI Agent node is an agentic step inside a deterministic workflow: the workflow is fixed, and the model reasons within that one step and returns a result. An agent is a standalone entity with its own identity, tools, triggers, sessions, and memory, which you publish and reuse. Both remain available.

If you can say the steps out loud, build a workflow. If the shape only appears at runtime, build an agent. Most real jobs are both: the agent decides whether, the workflow executes how.

Session memory is on by default, so conversations are preserved and you can review or continue them. Episodic memory across sessions is opt-in and requires an OpenAI credential. Agents do not learn or improve themselves automatically.

You can also ground an agent in your own content by uploading csv, pdf, markdown, or txt files to its knowledge base, which is available on n8n Cloud.

One turn with an agent counts as one execution. Workflow-tool calls and sub-agent calls do not count separately, and agents share the same execution quota as your workflows. High-frequency conversational use can consume executions faster than you expect, so it is worth watching at first.

Yes. Agents are generally available on n8n Cloud and self-hosted, and still ship behind a preview flag while the feature is under active development, with more landing release by release.

Some things are not here yet: always-on agents that watch event streams for hours or days, open-ended multi-agent orchestration beyond bounded sub-agent delegation, and external systems discovering or calling your agents.

No! n8n is built for both technical and non-technical users. With its intuitive drag-and-drop interface, you can create AI agents without writing code. However, developers can dive deeper using custom nodes and scripting when needed.

Yes. You can self-host to keep full control of your data, or run on n8n Cloud. Agent permissions use role-based access control (RBAC), credentials are scoped per tool, and sensitive tools can be set to pause for human approval before they run. Schedules, channels, and users always run against the published version, so draft edits never reach production. AI interactions can be configured to limit what data is sent to third-party models.

An assistant is someone you talk to. An n8n agent is something your company runs: published, versioned, bounded to the tools you gave it, with every session on the record. Beyond replying, it queries APIs, updates records, sends messages, and calls your workflows — and you can inspect each step it took.

AI Agents come with risks like hallucinations, runaway loops, or unintended actions. n8n mitigates this with:

• Step-by-step workflow control
• Rate-limiting and retries
• Memory limits
• Manual approval nodes
• Clear logging and debugging tools

You stay in control of what the agent can or cannot do.

Just sign up for a free trial on Cloud or explore our prebuilt templates to launch your first AI Agent in minutes. No setup headaches — just plug in your API key and go!