n8n-agents-official

n8n Agents

A first-class n8n Agent is a saved resource with its own instructions, model, tools, skills, sub-agents, memory, channels, scheduled tasks, versions, and spend limits. It is configured as JSON through the Agent MCP tools, never with Workflow SDK code. The AI Agent node inside a workflow is a different thing, covered by n8n-ai-nodes-official.

Agents are generally available. The UI shows a Preview tag because they change often and some features you might expect aren't there yet. They aren't ready for Enterprise yet: an Enterprise instance may have them enabled, but don't assume it. If create_agent isn't in your tool list, Agents aren't enabled on this instance. Say so and offer a workflow instead of guessing.

Non-negotiables

1. Always send a memory block when creating over MCP: "memory": { "enabled": true, "storage": "n8n" }. The UI adds it on every save. create_agent saves exactly what you send, and without it the Agent forgets the previous turn even inside one session (verified live).
2. Never publish without the user's explicit OK. A built Agent is a draft. Configuring a channel on an already published Agent connects it immediately, so confirm that too.
3. Publish what the Agent calls. Workflow tools need a When Executed by Another Workflow trigger, the same project, and a published workflow (validate_agent reports it). Saved sub-agents must be published too, and nothing checks that: validation and publishing pass, then production delegation fails and the parent quietly answers on its own.
4. Approval on anything that sends, changes, deletes, or spends, and know what the approver will see. See references/APPROVALS.md.
5. One mutate_agent per change, always with the latest configHash. A stale hash returns stale_config: call get_agent and retry.

Agent, workflow, or AI Agent node?

Decide before calling any build tool. The first question is the "say it out loud" test.

You can...	Build
Say every step before the request arrives, none needs judgment	Workflow
Say every step, one needs judgment (label, extract, summarize, write)	Workflow with an AI step: Text Classifier, Information Extractor, chain, or AI Agent node (n8n-ai-nodes-official)
Not say the steps until you see the request, and people talk to it or it acts on its own	First-class n8n Agent
Not say the steps, but one of the limits below applies	AI Agent node in a workflow
Not say the steps, and a fixed pipeline also needs the same brain	First-class Agent, called from the workflow with Message an Agent. The agent decides, the workflow acts

"Build me an agent" or "assistant" means a first-class Agent unless a limit applies. A Chat Trigger plus an AI Agent node is not a substitute for a requested Agent. If a request is really a fixed trigger plus fixed steps, explain the mismatch and ask before building a workflow instead.

Current limits that point to the AI Agent node

Need	Why an Agent can't do it today	Use instead
Custom approval message, or the approver must see the parameters in Slack	Approval is on/off with a fixed message. Channel cards show the tool name only	A workflow tool that runs its own send-and-wait approval with the values (can post in the same Slack thread)
Trigger or node data mapped exactly into a tool parameter	Tool expressions only see the model's own arguments (verified live for node and workflow tools)	AI Agent node, or Message an Agent with structured output and downstream nodes doing the mapping
One conversation shared by a native channel and workflows	Channel threads and Message an Agent session keys live in separate namespaces	A shell workflow on the platform's trigger, calling the Agent with Message an Agent
Approval inside a Message an Agent call	Fails with "Suspend/resume is not supported in workflow execution context"	Gate the action in the calling workflow

MCP build sequence

1. get_agent_builder_reference once. Its configSchema is the source of truth for field names; trust it over this skill when they differ.
2. search_projects, then discover_agent_assets (models, integrations, workflows, subagents, mcpServers), list_credentials, and search_nodes with usage: "agentTool" + get_node_types for node tools.
3. create_agent with model, credential, instructions, memory block, and ordinary tools. Skills, tasks, and custom tools come after, through mutate_agent, because they need server-generated ids.
4. mutate_agent, one operation per call. skill.upsert doesn't change configHash; config patches do.
5. validate_agent until valid: true and missing is empty.
6. One representative call_agent. It runs the draft through Preview with real tools, so side effects happen: if any tool without approval sends, writes, or pays, ask the user before the test. Every returned approval goes to the human; resume each with its continuation.
7. Report the draft with the url from validate_agent, plus the session link (see references/SESSIONS_AND_DEBUGGING.md), and ask whether to publish.
8. publish_agent only on an explicit yes. Channels: update_agent_integration (references/CHANNELS.md).

Details on every operation, error code, and version tool: references/MCP_BUILD.md.

Where depth lives

File	Read when
references/MCP_BUILD.md	Any mutate_agent operation, error codes, versions, revert, delete, availableInMCP
references/CONFIG.md	Writing or reviewing Agent JSON: every field, limits, defaults, what is versioned
references/TOOLS.md	Node, workflow, custom, MCP server, and built-in tools, input modes, what expressions can see
references/APPROVALS.md	Adding approval, what each surface shows, where approval can't happen, custom approval through a workflow tool
references/SKILLS.md	Adding Agent skills: how loading works, writing descriptions, references
references/SUB_AGENTS.md	Delegation, saved vs inline sub-agents, model routing by difficulty
references/KNOWLEDGE.md	Uploaded knowledge files vs external vector stores
references/MEMORY.md	Memory kinds, session scopes, session keys
references/CHANNELS.md	Slack, Telegram, Discord, Linear, n8n Chat, or "native channel or shell workflow?"
references/TASKS.md	Scheduled tasks
references/MESSAGE_AN_AGENT.md	Calling an Agent from a workflow
references/SESSIONS_AND_DEBUGGING.md	Anything went wrong, or the user asks where to see what the Agent did

For a chat bot built as a workflow shell (platforms without a native channel, slash commands, custom routing), the shell patterns live in n8n-ai-nodes-official references/CHAT_SURFACE_PATTERNS.md, section "Core as Message an Agent". Its one rule that can't wait: on surfaces that deliver the bot's own messages (Slack, Discord), a shell that posts replies must filter out the bot's user ID right after the trigger, or it triggers itself forever.

Writing instructions

Instructions are the system prompt. Same split as the node: persona, rules, output format, and inter-tool flow go in instructions; how to call a specific tool goes in that tool's description. Long procedures the Agent needs only sometimes belong in a skill, which costs tokens only when loaded. Say where to deliver results ("post the summary in #ops"), because a scheduled run has no one to reply to.

Anti-patterns

Anti-pattern	What goes wrong	Fix
create_agent without memory	Agent forgets the previous turn even in one session	Send { "enabled": true, "storage": "n8n" }
Chat Trigger + AI Agent node when the user asked for an agent	Loses channels, sessions, approvals, versions, tasks, and budgets	First-class Agent, unless a limit applies
Publishing because validation passed	The user didn't approve a live Agent, and channels go live with it	Report the draft and ask
Saved sub-agent left unpublished	Production delegation fails silently; the parent answers without it	Publish children before the parent goes live
Expression in a tool input that reads $('Webhook') or $json.<trigger field>	Only the model's arguments are visible, so the value is empty or wrong	Pass it through the message, or keep the mapping in a workflow
Approval on a tool the Agent calls under Message an Agent or a task	Run fails, or parks with no one to answer	Gate in the calling workflow; give tasks only safe tools
Same Message an Agent session key for two Agents in a project (e.g. both keyed on a bare Telegram id)	Keys are project-wide, not per Agent: the second Agent fails with "Failed to create the agent execution record"	Prefix keys per Agent: support-<telegramId>
Skill body says "read references/x.md" vaguely	The model often skips it	"Load references/x.md with load_skill (filePath) before answering"
Integrations set through config.patch	Rejected	update_agent_integration
Slack connected with a credential built for a Slack Trigger	Passes every check, receives nothing	Managed setup, or an app whose request URL points at this Agent