n8n Agents vs Workflows

Thirty controlled support turns produced the same 20 completed cases in both designs, but the new n8n Agent needed 40 scoped workflow calls and preserved 20 conversation sessions. In n8n Agents vs Workflows, the decision is not which one automates more: keep a workflow in charge when the sequence is fixed, and use an Agent only when a conversation must decide the next step.

n8n Agents vs Workflows: The Verdict

Pick an n8n workflow when you can draw the sequence before the run starts. Pick an n8n Agent when the next useful action depends on what a person says, what a tool returns, or what the conversation has already established. For production support and operations, the strongest design is usually hybrid: the Agent chooses, while narrow workflows do the work.

That distinction matters more than whether a model appears somewhere in the automation. A workflow can call a model and remain deterministic. An Agent can call workflows and remain agentic because the model, not the canvas, chooses the next tool.

The deciding question is simple: who should own the next step? If the answer is the automation designer, keep a workflow in charge. If the answer is the model after reading the current conversation and tool results, use an Agent. n8n makes the same recommendation in its launch material: fixed sequences fit workflows, while open-ended requests fit Agents that can work out the steps. n8n’s launch explanation is unusually direct about that boundary.

What Changed on September 25

n8n introduced the new Agents surface on September 25, 2026. An Agent is now a first-class project artifact with its own model, instructions, tools, memory, sessions, draft, published version, channels, and schedules. It sits beside workflows rather than inside one particular canvas. The current n8n Agents documentation describes it as the place for work that is too open-ended for a fixed workflow.

That shared identity is the useful change. The same published Agent can answer in a channel, run on a schedule, or receive a message from another workflow. Its session history records the conversation, tools, outputs, errors, and pending approvals. Editing the draft does not silently change the published version.

The model still should not receive broad access to every system. A workflow tool gives it a narrow action contract: named inputs, a controlled credential, a known output, and an optional approval boundary. One early practitioner reaction captured the point well: reusing existing workflows as tools is more useful than adding another chat wrapper.

n8n AI Agent Workflow Builder: What Changed

The old pattern assembled an agent inside a workflow from a Chat Trigger, memory, an AI Agent node, model, and tools. That pattern still works. The new builder moves the durable agent identity, sessions, versions, and multiple entry points into a shared artifact.

This is not a cosmetic editor change. It changes ownership:

• The Agent owns the conversation and tool-selection loop.
• Published workflows own bounded actions such as looking up an account or drafting a reply.
• The calling channel, schedule, or workflow owns when the Agent receives work.
• An approver owns the final decision for a sensitive tool.

n8n Agents vs AI Agent Node

The existing AI Agent node is still a workflow node. It connects to a chat model and at least one tool, then chooses tools during that workflow run. n8n says existing AI Agent node builds continue to work. The live node documentation still describes that model-plus-tools design.

Use the node when the agent belongs to one workflow and that workflow should own its trigger, memory arrangement, and lifecycle. Use the new Agent when one identity must persist across conversations or be called from several places. Migrating is optional, not a prerequisite for keeping an existing node-based agent running.

The Same Support Task, Built Both Ways

The controlled comparison used a disposable local n8n 2.40.7 instance, one deterministic OpenAI-compatible local model endpoint, and one fixed JSON fixture. It was a routing and orchestration test, not a model-quality or cloud-latency benchmark.

n8n AI Agent Example: Support Triage

The fixture contained 20 synthetic support tickets:

• Ten arrived with a fixed format containing ticket ID, account ID, product area, and issue.
• Ten intentionally omitted account ID and product area, so a useful system had to ask one follow-up question.
• Every final case had a fixed expected queue: billing, technical, or general.

Each design processed the same 30 user turns. The ten complete tickets took one turn each. The ten incomplete tickets took an initial question and one follow-up, for 20 more turns.

1. 1

   Build the fixed sequence

   The workflow used three nodes: a webhook received the ticket, the shared local model classified it, and a parser returned structured JSON. Every incoming message followed those nodes in that order.

2. 2

   Build the Agent

   The Agent used the same model, explicit triage instructions, stored session memory, and three published workflow tools: Get Account Context, Draft Support Reply, and Page On Call. The first two were available directly. Page On Call required approval.

3. 3

   Score complete work

   A case counted as complete only after the output contained the expected final queue. Clarification questions, executions, sessions, model endpoint requests, workflow-tool calls, failures, and unnecessary actions were logged separately.

Both designs routed all 20 final cases to the expected queue. The result does not prove that Agents and workflows have equal judgment. The stub was deliberately deterministic so the test could isolate orchestration. It shows where the extra machinery appears: the fixed workflow made one model request per turn, while the Agent called the model repeatedly to choose tools, consume tool results, produce the response, and maintain its run.

The Agent’s 100 endpoint requests consisted of 70 streaming reasoning or tool-loop requests and 30 non-streaming helper requests in this local setup. Treat that as a warning to meter provider usage, not as a universal multiplier. A different model, memory configuration, prompt, or product revision can change the count.

When to Use n8n Agents

Use an Agent when a conversation changes the plan. A support request that begins as “billing is wrong” may need an account lookup, a clarifying question, a policy check, or an approval before any action. You cannot know the correct branch until the missing information arrives.

Use a workflow when the plan is already known. A nightly export, a webhook-to-CRM sync, or a lead-enrichment sequence benefits from explicit nodes, predictable retries, and an execution path you can inspect without reconstructing a model’s decision.

Control and debugging winner: workflow

A workflow is easier to reason about because the possible next nodes are on the canvas. If a deterministic process fails, the execution tells you which node failed. That is the right default for money movement, record deletion, entitlement changes, and other operations where flexibility is more liability than benefit.

An Agent session gives you a different kind of trace: messages, tool choices, outputs, errors, and approvals. That is useful, but it does not turn a probabilistic decision into a declared graph. If the sequence never needs to vary, adding a reasoning loop only creates another failure surface.

Conversation and state winner: Agent

The Agent wins when the job spans several turns. Sessions are stored and resumable, and session memory is on by default. In the fixture, ten incomplete tickets stayed inside their original session when the missing account and product details arrived. The fixed workflow produced the same answer only because the follow-up message restated enough context for a stateless classifier.

That difference gets larger in real support work. If a second message says only “the EU account,” a workflow must reload the earlier ticket from a data store or receive the history in its input. An Agent session already has the conversation context. Episodic memory can reach across sessions too, although n8n currently requires an OpenAI credential for that feature.

Sensitive-action winner: workflow, with an Agent in front

The safest hybrid gives the Agent read-only tools freely and puts approval on side effects. In a separate urgent-ticket smoke test, the Agent completed its read-only account lookup, selected Page On Call, and then suspended. The paging workflow did not run until an approver would accept the tool call.

That is the right security model: an Agent can recommend or request an action, but a narrow workflow and explicit approval own the blast radius. Credentials attach to the tool, so the Agent does not need one broad credential that can do everything.

Overall winner: hybrid

The new surface is strongest as a conversational control plane over workflows, not as a replacement for them. The Agent interprets, asks, and chooses. Workflows validate, mutate systems, and return structured results. This division also lets you test actions independently from the model that selects them.

n8n Agents Execution Cost

As of September 26, 2026, n8n does not price Agents on a separate plan. One Agent turn counts as one execution, and Agent plus workflow executions draw from the same quota. n8n’s launch post adds an important detail: workflow-tool and sub-agent calls inside that turn do not count as separate plan executions.

The current annual-billing prices are $20 per month for Starter with 2,500 executions and $50 per month for Pro with 10,000 executions. Both prices were verified against the live n8n pricing page, which shows a 17% saving on the annual toggle. For the broader plan tradeoffs, see the separate n8n pricing analysis.

Consider the planning scenario from the brief: 200 conversations at three turns each.

• 600 Agent executions come from 200 × 3. Internal workflow tools remain inside those turns for quota purposes.
• On Starter, $20 ÷ 2,500 is $0.008 of subscription quota per included execution. The 600 turns allocate $4.80 of that monthly subscription and leave 1,900 executions.
• On Pro, $50 ÷ 10,000 is $0.005 of subscription quota per included execution. The same 600 turns allocate $3.00 and leave 9,400 executions.

Those quotients are allocation math, not marginal invoice rates. If the 601st turn still sits inside the plan allowance, it does not add an $0.008 line item to the Starter bill.

The crossover is a quota boundary, not an Agent-versus-workflow discount. Starter fits 833 complete three-turn conversations at 2,499 executions; conversation 834 reaches 2,502 and crosses the 2,500 allowance. Pro fits 3,333 such conversations at 9,999 executions; conversation 3,334 reaches 10,002 and crosses the 10,000 allowance.

If a fixed workflow receives one webhook per chat message, it also consumes 600 executions for the same 600 turns. The execution bill is therefore a tie. The fixed workflow can still be cheaper at the model layer because it may make one model request where an Agent makes several decisions around tools.

Model usage sits outside the execution quota. n8n Gateway credits use a separate prepaid balance, and you can choose your own provider credential instead. Gateway-credit documentation says a zero balance makes supported nodes fail until the owner tops up, enables auto top-up, or switches credentials. The local 30-versus-100 request result is why both execution count and model spend belong on your dashboard.

Sessions, Versions, Approvals, and Workflow Calls

The new Agent earns its place when several entry points need the same behavior. n8n stores every conversation as a session, including messages, tools, and pending approvals. The measured fixture created 20 sessions, one per ticket, and the ten follow-up tickets each resumed the existing session for their second turn.

Versioning separates experimentation from production. The draft saves as you edit, while Publish takes a snapshot that channels, schedules, and production chats use. In the local check, a draft-only instruction produced a new draft version while the active version ID and published instructions stayed unchanged. That is the behavior you want when refining a prompt during live operations.

n8n Message an Agent: One Agent, More Than One Entry Point

The Message an Agent node lets a workflow call an existing published Agent. A two-node smoke workflow sent a billing ticket to the same Support Triage Agent, received a complete result, and recorded the same two scoped tool calls: account context, then reply drafting. The node also accepted a custom session key, so a workflow can continue a known conversation instead of starting from zero.

This creates a useful composition pattern:

1. A deterministic workflow receives and validates an event.
2. Message an Agent passes only the conversational decision to the published Agent.
3. The Agent chooses from scoped workflow tools.
4. The calling workflow receives the Agent’s text, usage, tool-call log, and session reference.

Avoid the circular version of this architecture. A workflow that calls an Agent should not also be attached to that Agent as a tool. Keep entry-point workflows and action workflows separate, with names that make the boundary obvious.

Approvals fit inside the same session trace. When a sensitive tool is selected, the Agent pauses and presents its arguments. Approve lets it continue from that point; Reject cancels the action. That is more useful than a generic “human in the loop” promise because the approver can see the proposed tool and input before the credential is used.

For high-risk automations, session history is still not enough by itself. Send Agent failures and suspicious tool choices into an independent review path, and keep retention, redaction, and alerting aligned with the data involved. The AI agent failure-analysis guide covers that separate observability layer.

Switching Without Rebuilding Everything

Switching should mean wrapping stable workflows, not redrawing them inside a prompt. The existing workflows already contain the valuable parts: credentials, validation, API calls, transformations, and failure handling.

1. 1

   Keep the deterministic spine

   Leave schedules, webhooks, validation, and irreversible writes in workflows. Do not migrate a fixed sequence simply to make the architecture look agentic.

2. 2

   Turn actions into contracts

   Give each callable workflow a narrow input schema and a structured output. Separate read-only lookups from side effects so approval can be applied only where it is needed.

3. 3

   Attach the smallest tool set

   Start the Agent with the few workflows needed for one job. Specific names and descriptions help the model choose correctly and make the session log easier to audit.

4. 4

   Test conversations, not prompts

   Use fixed-format cases, missing-information cases, repeated messages, and unsafe requests. Score final outputs, clarification turns, tool calls, rejected actions, and model usage before publishing the snapshot.

5. 5

   Add entry points last

   Once the published Agent is stable, connect a channel, schedule, or Message an Agent workflow. Reuse the same identity rather than cloning instructions across several canvases.

Do not switch if your sequence is fixed, your current AI Agent node belongs to one workflow, or your compliance model cannot accept preview software. Self-hosted Enterprise and queue-mode deployments have an even clearer answer today: wait. If you are choosing the larger automation stack around n8n, the AI automation tools comparison gives that broader context.

Migration cost hides in interfaces, not node count. Every workflow tool needs clear inputs, controlled credentials, predictable outputs, duplicate-safe actions, and an owner for approvals. An Agent exposes weak contracts quickly because it may call the same tool in an order the original workflow designer did not anticipate.

Preview Limits and the Recommendation

n8n Agents are in Preview. On n8n Cloud, they are available to everyone running the latest stable version. Self-hosting starts at 2.32.3 and requires enabling the agents module for a manual setup. The full AI-assisted builder is optional, but self-hosted knowledge bases need a Daytona sandbox and channels need a public webhook URL.

Two limits can stop an otherwise attractive migration: Agents are not ready for self-hosted Enterprise, and queue mode is not supported. n8n also warns that self-hosted channel connections such as Telegram can fail, so regular mode is the current recommendation.

The honest production architecture is conservative. Keep the workflow in charge of every sequence you already know. Put the Agent only at the point where a conversation must choose what comes next. Give it narrow workflows as tools, store the session, publish a tested snapshot, and put approval in front of every meaningful side effect.

That is enough autonomy to make the new surface useful without asking a preview Agent to become your entire automation system.

Frequently Asked Questions

Can I use n8n to build agentic AI?

Yes. The new Agents surface builds a persistent Agent that can choose tools across sessions, while the existing AI Agent node builds agentic behavior inside a workflow. Pick the scope that matches the job.

What are the big 4 AI agents?

There is no authoritative “big four” for AI agents. Choose by the work, required integrations, approval model, deployment constraints, and cost rather than a generic popularity list.

What are the 5 types of agent in AI?

There is no universal five-type taxonomy. For an n8n build, the useful distinction is whether a fixed graph owns the next action or a model chooses among bounded tools.

What is the difference between n8n workflows and agentic workflows?

An n8n workflow follows a declared graph of nodes. An agentic design lets a model choose the next action, while workflows can still execute the approved tools underneath it.

Is ChatGPT an agentic AI?

A chat response is not inherently agentic. Agentic behavior means pursuing a goal by selecting actions or tools, observing results, and deciding what to do next.

What are the four types of agents?

There is no single four-type standard that governs n8n. Evaluate operational traits instead: state, planning, tool access, and autonomy.

What are the top 3 AI agents?

There is no universal top three. The right Agent depends on the systems it must access, how much control it needs, and where it can run.

What are the 7 types of AI?

Seven-type lists are educational taxonomies, not an architecture rule. They do not decide whether your support or operations process belongs in an n8n workflow or Agent.

What are the 5 parts of an AI agent?

For an n8n implementation, start with model, instructions, tools, memory, and access controls. Add knowledge, channels, schedules, or sub-agents only when the use case requires them.

If the conversational decision point is the hard part, I can help design and build the Agent while keeping your workflows as the control surface.