n8n AI Agent Design

n8n Agents

The n8n AI Agent node (@n8n/n8n-nodes-langchain.agent) is a multi-turn LLM driver with sub-nodes for the model, memory, tools, and an optional output parser. This skill is the deep guide to designing agents and the LangChain family around them. For the high-level "where an agent fits in a workflow" picture, see n8n-workflow-patterns ai_agent_workflow.md — this skill goes one level down into how to build it well.

For node-type formats: in workflow JSON the LangChain nodes use the long @n8n/n8n-nodes-langchain.* form (.agent, .lmChatOpenAi, .memoryBufferWindow, .outputParserStructured, .toolWorkflow, .toolHttpRequest, .toolCode). When you call get_node / validate_node, use the short form (nodes-langchain.agent). See n8n-mcp-tools-expert for the format rules.

Pick the right node first

Reaching for an Agent when the task is one-shot classification or extraction is the most common over-build. Decide before you wire anything:

Text Classifier detail (the Agent + Switch anti-pattern): every category needs both a name AND a description. The model routes against the description, not the name — a category with no description gets picked by coin-flip. Set options.enableAutoFixing: true for robustness on edge inputs. One node, N branches, done. Reaching for an Agent that "decides" then a Switch that "routes" is two nodes plus prompt boilerplate for what Text Classifier does natively.

Chat-model nodes (.lmChatOpenAi, .lmChatAnthropic, .lmChatOpenRouter, …) are sub-nodes — they don't run standalone. They wire into a chain, agent, classifier, or extractor via the ai_languageModel connection.

The sub-node pattern

The Agent has a main input (the prompt / user message) and up to four sub-node slots, each wired by its own ai_* connection type:

A sub-node connects FROM itself TO the agent. In workflow JSON the connection lives on the sub-node, keyed by the ai_* type:

"Main LLM": {
  "ai_languageModel": [[{ "node": "AI Agent", "type": "ai_languageModel", "index": 0 }]]
},
"Simple Memory": {
  "ai_memory": [[{ "node": "AI Agent", "type": "ai_memory", "index": 0 }]]
},
"Search customer DB": {
  "ai_tool": [[{ "node": "AI Agent", "type": "ai_tool", "index": 0 }]]
}

Multiple tools all connect into the same ai_tool index 0 — they stack, they don't fan into separate indices. With n8n_update_partial_workflow you wire each with an addConnection op using sourceOutput: "ai_tool". The agent puts its final answer in $json.output (not .text, not .response) — downstream nodes read {{ $json.output }}.

See EXAMPLES.md for a complete stateless agent-core node-object snippet.

Two non-negotiables

1. Tool names and descriptions ARE part of the prompt. The model picks a tool by reading its name and description — nothing else. A tool named tool1 with an empty description is invisible to the model: it skips it, mis-selects it, or hallucinates parameters. There's usually no error — just an agent that "won't use my tool". Treat both like API design. → TOOLS.md
2. Structured output must parse AND autoFix. An outputParserStructured with autoFix: true and a coding-capable fixer model is the production pattern. Without autoFix, one malformed JSON response halts the whole workflow. → STRUCTURED_OUTPUT.md

Strong defaults

• Per-tool usage goes in the tool description, not the system prompt. Anything about how to call this specific tool belongs with the tool, so it travels across agents and keeps the system prompt focused. → SYSTEM_PROMPT.md
• Sub-workflow tools (.toolWorkflow) for anything multi-step. Any workflow becomes a tool with typed $fromAI() inputs, and composes with branching, error handling, and reuse. Default here when in doubt. → SUBWORKFLOW_AS_TOOL.md and n8n-subworkflows.
• Wrap tools with user-visible side effects in human review. Sends, payments, refunds, account changes get gated behind an approval node so a human signs off before the tool fires. → HUMAN_REVIEW.md
• Raise maxIterations. The default tool-call cap is low (single digits on most versions) — fine for a one-tool agent, far too low for a multi-tool agent that chains several calls per turn. It surfaces as "max iterations reached" or empty output. Set options.maxIterations to a realistic ceiling (15 for a focused sub-agent, 50-200 for a broad orchestrator).
• Put the current date in the system prompt via {{ $now }} (or {{ $now.format('DDDD') }}). A hardcoded date is stale immediately.

The four tool types

Pick the lightest option that covers the job:

There is also a Custom Code Tool (.toolCode) for pure inline computation — but its runtime contract (string in / string out, no $fromAI, no $helpers) is owned by the n8n-code-tool skill. Read that before writing one. Rule of thumb: if you find yourself reaching for $fromAI() inside the code, you want .toolWorkflow instead.

$fromAI(): how the agent fills tool parameters

Tool parameters the agent should decide are wrapped in $fromAI(). It is a real n8n expression helper, used inside a tool node's parameter expressions:

={{ $fromAI('paramName', 'what to put here — be specific: format, range, example', 'string') }}

• paramName — the name the model uses internally (snake_case or camelCase, be consistent).
• description — tells the model what value to produce. It is part of the prompt — write it like JSDoc.
• type (optional) — 'string' (default), 'number', 'boolean', 'json'. A wrong-typed value fails the call.
• defaultValue (optional) — used when the model omits it.

$fromAI() carries JSON only — it cannot carry binary (no base64, no file bytes). And not every parameter has to be $fromAI: plumb identity, authority limits, and correlation IDs (userId, refund caps, sessionId) deterministically from workflow context so the agent can't get them wrong or even see them. → TOOLS.md for the full anatomy and the "give the agent a button, not a steering wheel" pattern.

System prompt vs tool description

Why split it: a well-described tool works in any agent that drops it in, tool details only "load" when the model considers that tool (token efficiency), and you update one tool description instead of a paragraph buried in a 5000-token prompt. → SYSTEM_PROMPT.md

Structured output: when and how

Add an outputParserStructured sub-node (wired ai_outputParser) when downstream needs strict JSON, not free-form text. Two rules:

1. Use schemaType: 'manual' with a real JSON Schema, not jsonSchemaExample. An example can't express required-vs-optional, enums, numeric ranges, or array constraints — you outgrow it the first time the shape gets non-trivial. Reach for fromJson + an example only for throwaway shapes.
2. autoFix: true with a coding-capable fixer model. Wire a second model into the parser's ai_languageModel slot. Reconciling broken JSON against a schema is a coding task — a weak fixer just produces another malformed retry and burns tokens.

→ STRUCTURED_OUTPUT.md for the schema patterns, the load-bearing "DO NOT wrap in markdown" retry line, and the parse-failure cookbook.

Memory: brief mental model

Memory is a sub-node (ai_memory). Without it, every call is stateless — correct for one-shot tasks (classify, summarize). With it, the agent holds a conversation, keyed by whatever expression you bind to sessionKey.

• memoryBufferWindow — keeps the last N exchanges per key and persists across executions via n8n's store. The default for chat. contextWindowLength defaults to 5, which is very low — 50 is a saner starting point. Messages past the window are gone entirely.
• memoryPostgresChat / memoryRedisChat — only when memory must be read outside the agent (your own UI, analytics, cross-system). Not needed just to survive restarts; BufferWindow already does that.

Plumb a stable key from the trigger to memory consistently. Chat triggers fill sessionId automatically; for other surfaces derive one (Slack thread_ts, a webhook conversation ID). Never hardcode sessionId: 'default' and never put sessionId behind $fromAI (the model will fabricate a UUID). → MEMORY.md

Binary and the agent boundary

This is the seam that trips people up:

• The model CAN see uploaded images (vision) via options.passthroughBinaryImages: true on the agent.
• Tools CANNOT receive binary. $fromAI() is JSON-only — no base64, no bytes, even through non-AI bindings.
• The agent's output is text-shaped (or structured-text with a parser). When a model returns image/audio/video bytes, the Agent doesn't surface them at all — there's nothing to recover downstream.

Workaround: pre-stage uploads to storage before the agent runs, inject the storage keys into the system prompt, and let tools accept the key as a string parameter and re-fetch internally. For one-shot media generation, skip the agent and call the provider's native single-call node directly.

The binary mechanics (which storage, how to stage, how to re-fetch) are owned by n8n-binary-and-data — see its agent-tool binary reference. This skill only marks the boundary; don't re-derive the mechanics here.

Human review (gate destructive tools)

When a tool's effect needs human sign-off before execution (sends, payments, refunds, account changes), wrap it with a review tool node — slackHitlTool, discordHitlTool, telegramHitlTool, gmailHitlTool, etc. (n8n names these "Hitl" / human-in-the-loop). The review node sits between the wrapped tool and the agent on the ai_tool connection: wrapped tool → review node → Agent.

Whether sign-off is needed is a product/policy call — surface the question to the user, recommend based on blast radius, and let them decide.

The critical rule: show the actual parameters the wrapped tool will receive. Use the literal {{ $tool.parameters.<name> }} in the approval message, never a $fromAI() paraphrase — otherwise the human approves text the model made up, not the call about to fire. → HUMAN_REVIEW.md

Chat agents (Slack, Discord, Teams, Telegram)

The one non-negotiable, regardless of complexity: any chat-triggered workflow that posts a reply MUST filter out the bot's own user ID, or its own replies re-trigger it in an infinite loop that burns runs and tokens. Prefer trigger-level filtering when available (Slack Trigger's options.userIds is an exclusion list — put the bot ID there); otherwise filter $json.user !== '<BOT_USER_ID>' in the first node after the trigger.

Beyond the filter, a simple bot (trigger → agent → reply) lives fine in one workflow. Split into shell + core + sub-agents only once you need loading UX, sub-agents, multi-surface reuse, or robust error handling:

• Shell — trigger, anti-loop filter, event-type Switch, loading/error UX, renders the reply. No LLM.
• Core — stateless agent, chatInput + threadId inputs, memory keyed on threadId, tools and sub-agents.
• Sub-agents — one narrow domain each, called via .toolWorkflow, stateless (full context in chatInput).

→ CHAT_AGENT_PATTERNS.md for per-surface semantics, threading-as-session, and the full topology.