GoHighLevel Conversation AI Flow Builder V2 + Subagents

GoHighLevel just gave Conversation AI a brand new Flow Based Builder V2, and your managed agents can now split a job into a team of subagents.

Flow Based Builder V2 is a Labs Conversation AI canvas that builds inbound chat agents as nodes, capture fields, tools, routers and end steps, while Agent Studio managed agents can create specialist subagents that only get the tools they need. This guide is the written version of the Faceless GHL tour 👇

New to the product? Start with What Is GoHighLevel or the Explained tour. Already building prompt agents? See Conversation AI Build Agents with AI. Mapping automations after? Read Workflow Maps Explained.

TL;DR: Flow Builder V2 + Subagents

• Enable: Labs → Conversation AI Flow Based Builder V2, then Create Agent → Flow Based Builder.
• Build: use Build with AI, answer its follow-up questions, then edit nodes on the canvas.
• Test: pick a contact, chat, and read the execution timeline under every reply.
• Subagents: in Agent Studio, describe a managed job in one sentence and ask it to split into specialists.
• Limits: flow builder is inbound only; outbound still needs a prompt agent; research subagents stay read only.

What Flow Builder V2 and subagents are

Until now, most Conversation AI agents in HighLevel were prompt based: one big instruction block, channel settings, and tools. That still works for many bots. Flow Based Builder V2 adds a second path. You build the agent as a visual flow of nodes, starting with Start Chat, an AI Agent and an End Node, then adding capture steps, routers, waits, API calls, appointment booking and more.

Subagents are not the same feature as the flow canvas. They live under managed agents in Agent Studio. You describe a recurring or multi-step job, ask HighLevel to split it, and it creates named specialists such as an Opportunity Reviewer and a Follow Up Task Creator. The main agent plans. Each specialist runs with a smaller tool set.

If you already ship prompt agents from the Build your Agents with AI flow, keep that for simple briefs. Reach for Flow Builder when ordered capture, branches and tool steps keep going wrong inside one prompt.

Think of the two features as different zoom levels. Flow Builder V2 is the inbound receptionist that greets a stranger on chat, collects details in order, and books a slot. Subagents are the back-office team that runs on a schedule, checks stalled opportunities, and creates tasks for humans. Both are AI agents in HighLevel. They solve different jobs and live in different menus.

Turn this into a system you can sell. Watch the free masterclass for the agency method, or get free access to the community through Greg's GoHighLevel link.

Who this is for

Agencies building front desk bots

Plumbers, dentists, gyms and home services need the same pattern: collect name and phone, answer service questions, book a visit, push emergencies to the earliest slot. Flow Builder V2 is built for that ordered inbound path.

Operators tired of one giant prompt

When a prompt agent keeps skipping capture or booking the wrong calendar, a canvas with separate nodes is easier to fix. You click the node, change the instructions, and retest.

Teams with weekly CRM follow-up jobs

Managed agent subagents suit scheduled work such as finding stalled opportunities, reading recent conversations and creating tasks for owners. One sentence can produce the planner plus specialists.

Builders who need a clear debug trail

The execution timeline under each test reply is the quiet hero of Flow Builder V2. When a client says the bot skipped the phone number or never offered a booking, you open Test, replay the chat, and read which node ran. That beats guessing inside a three-page prompt.

If you are still learning where Conversation AI sits in the wider menu, keep What Is GoHighLevel open in another tab. Flow Builder does not replace CRM, calendars or workflows, it sits on top of them and calls those tools when the graph says so.

Watch the walkthrough

Prefer to watch first? The Labs → canvas → Build with AI → test timeline → Agent Studio subagents tour is below, recorded in a Northline Plumbing demo sub account.

• 0:00 GoHighLevel Conversation AI flow builder and subagents
• 0:16 The plan
• 0:32 Turn on Flow Based Builder V2 in Labs
• 0:48 The new canvas and nodes
• 1:13 Build with AI from one prompt
• 1:43 The flow it built
• 1:59 Test chat and the execution timeline
• 2:16 Managed agents: build from one sentence
• 2:34 Subagents: two specialists
• 2:53 Test mode run
• 3:10 Tips before you use it
• 3:37 More GoHighLevel updates

The plan

The video plan is simple. First switch on the new flow builder and get AI to build a front desk bot for the plumbing account. Then test it with a contact and see exactly what it did. After that, build a managed agent from one sentence, show the subagents it creates, and finish with tips so you do not get caught out.

That same order is what this article follows, so you can pause the video at any chapter and read the matching section.

Turn on Flow Based Builder V2 in Labs

Open your sub account (Northline Plumbing and Heating in the demo). Go to Labs and find Conversation AI Flow Based Builder V2. Switch it on.

Then go to Conversation AI and click Create Agent. You now get two options: a Prompt Based Agent, or the Flow Based Builder. Pick the flow one.

If you only see the old prompt path, the Labs flag is still off at agency or account level, or the feature has not rolled to your location yet. Flip the Labs switch, refresh, and try Create Agent again.

Treat Labs like a feature flag you own. Turn Flow Based Builder V2 on for the agency when you are ready to train your team on the canvas. Leave it off if your staff should only ship prompt agents for now. The Create Agent screen is the quick check: two options means the flag is live; one option means it is not.

Demo tip from the Northline account: switch Labs on first, then open Conversation AI. Creating the agent before the flag is on just sends you back into the familiar prompt builder, which is fine, but it is not the V2 canvas this guide covers.

The new canvas and nodes

The canvas starts with three nodes: Start Chat, an AI Agent, and an End Node. Across the top you get Build, Deploy, Variables, Settings and Test.

Click the plus and you can add a wide node palette. In the walkthrough that includes AI Agents, a Global AI Agent, Router, Wait, Bot Transfer, Search the web, MCP Server, API Call, Appointment Booking, and capture nodes for email, phone and text.

You can drag those in by hand. Most people will not. The useful path is Build with AI on the left, which drafts the graph from a plain English brief, then lets you edit any node.

Still, knowing the palette matters. Capture nodes force order: name before phone, phone before booking. Routers branch emergencies away from normal repairs. Wait and Bot Transfer cover handoffs. API Call and MCP Server are there when you need systems outside HighLevel. Appointment Booking only works once a calendar is attached, the warning badge is not optional decoration.

Variables and Settings at the top are where you keep shared values and agent-level options without burying them inside a single prompt. Deploy is the go-live control after Test looks clean. Build is where you stay while the graph is still changing.

Build with AI from one prompt

On the left, open Build with AI and type what you want. In the demo the brief was:

Build a front desk bot for Northline Plumbing and Heating, collect the customer's name and phone number, answer questions about our services, book a repair visit, and send emergencies like a burst pipe straight to booking.

That one sentence covers capture, knowledge answers, booking and an emergency fast path. You do not need to name every node. The builder proposes the graph, then asks clarifying questions before it applies changes.

Turn this into a system you can sell. Watch the free masterclass for the agency method, or get free access to the community through Greg's GoHighLevel link.

It asks questions back

The part worth pausing on: Build with AI asks questions instead of guessing. In the tour it wanted to know whether a knowledge base existed and which calendar to use.

The answers in the demo were: no knowledge base yet, so use web search on services, and emergencies get the earliest available slot. That is better than silent assumptions that put every lead on the wrong calendar.

When you run Build with AI on your own account, answer those prompts carefully. Wrong calendar or missing knowledge source is the usual reason a pretty graph fails in Test.

A good answer pattern looks like this: name the calendar the front desk should use, say whether web search is allowed when the knowledge base is empty, and define what counts as an emergency in one short line. Burst pipe, no heat in winter, and flooding are examples from a plumbing niche. Your niche will differ, write the rule once so the builder does not invent one.

If Build with AI asks something you do not know yet, stop and connect the calendar or upload a thin services FAQ before you accept the graph. Applying a half-answered draft just creates a prettier version of the same gap.

The flow it built

After the answers, the builder applied the graph to the canvas. It added a sequential node that captures name and phone first, then an AI agent called Northline Front Desk with its own prompt, plus web search and appointment booking as tools.

Click any node to change the instructions yourself. That is the edit loop: AI drafts the structure, you tighten the wording, reconnect tools, and keep the capture order you actually want.

Test chat and the execution timeline

Click Test, pick a contact, and start a chat. In the demo the opener was “my heating has stopped working.” The bot asked for a name, then a phone number.

Under every reply there is an execution timeline. You can see which step ran and how long it took. That is the debug feature prompt-only agents never gave you clearly: when a bot goes wrong, you stop guessing which instruction fired and read the timeline instead.

Use Test on a throwaway contact before you Deploy. Confirm capture order, tool calls and the end path. If appointment booking shows a warning, fix the calendar before you chase prompt wording.

Read the timeline like a flight recorder. You want to see capture nodes fire before the agent answers service questions, and you want appointment booking to appear only after the required fields exist. If the timeline shows the agent jumping ahead, tighten that node or add a capture step, do not only append more prose to the agent prompt.

When the chat feels right, repeat Test with an emergency phrasing and a normal repair phrasing. The Northline brief asked for burst-pipe style urgency to go straight to booking. Your test script should prove both paths, not only the happy path.

Managed agents from one sentence

The second half of the update is not on the Conversation AI canvas. Go to Agent Studio, click the plus to build a managed agent, and describe the job in one sentence.

Every Monday, find opportunities that haven't moved in fourteen days, check their recent conversations and create follow up tasks for the owners. Split this into subagents.

That last line matters. Without “split this into subagents,” you may get a single managed agent. With it, HighLevel is nudged to create specialists under the planner.

Managed agents suit scheduled CRM work. Conversation AI flow agents suit inbound chat. Keep those jobs separate in your mental model even though both are “agents” in the product.

Write the managed-agent sentence the way you would brief a junior ops hire. Include the cadence (every Monday), the filter (fourteen days with no move), the check (recent conversations), the output (follow up tasks for owners), and the split instruction. Vague briefs create vague specialists.

After the agent appears, open the Subagents row and confirm each specialist’s tools. If a reviewer has write tools it should not have, fix that before Test Mode. The whole point of the split is smaller tool kits, not three copies of the same overloaded agent.

Subagents: two specialists

In the walkthrough, HighLevel built an agent called Opportunity Follow Up that runs on a schedule. Under Subagents it split out an Opportunity Reviewer to do the checking. On another run it also split out a Follow Up Task Creator.

The main agent plans the job. Each specialist only gets the tools it needs. HighLevel's own tests say that pattern is about half the cost and noticeably faster than one overloaded agent doing every step.

That split also matches the tip later: research-style subagents stay read only. Keep write work on a separate specialist so a reviewer cannot quietly edit CRM records.

Test mode run

Click Test Mode, pick a contact and click Start. In the demo, in under a minute the agent checked one hundred and twenty opportunities across all four pipelines, found none stalled yet, changed nothing, and reported that none of them had an owner, so tasks would have nowhere to go.

That owner catch is the useful part. A green “ran fine” without owners would have created a false sense of readiness. Test Mode surfaced a data problem before the Monday schedule ever ran.

When you test your own managed agent, watch for the same class of findings: missing owners, empty pipelines, or permissions that block task creation. Fix the CRM shape before you trust the schedule.

The Northline run checked one hundred and twenty opportunities across four pipelines in under a minute and changed nothing because nothing was stalled. That is a successful test even though no tasks were created. A useful managed agent is allowed to report “all clear”, what you cannot skip is the owner warning. Without owners, follow-up tasks have nowhere to land.

Before you enable the real Monday schedule, assign owners on active opportunities or add a step that routes unowned deals to a default user. Then run Test Mode again and confirm tasks would land on a real person.

Tips before you use it

1. Connect a calendar when booking warns

If appointment booking has a little warning on it, connect a calendar. The first test booking in the video failed for exactly that reason. Fix the warning before you rewrite prompts.

2. Flow builder is inbound only

The new flow builder is for inbound chats only. Outbound messages still need a prompt-based agent. If your use case is outreach or follow-up SMS blasts, stay on the prompt path or a workflow, not Flow Builder V2.

3. Research subagents are read only

Research subagents can check data but cannot change your records. Pair them with a task-creating specialist when the job needs writes. Do not expect a reviewer-only subagent to update opportunities.

4. You do not need flows for every simple bot

If one brief is enough, use a prompt-based agent. Use flows when one big prompt keeps getting steps wrong and you need ordered capture, tools and branches on a canvas.

Turn this into a system you can sell. Watch the free masterclass for the agency method, or get free access to the community through Greg's GoHighLevel link.

Prompt vs flow agents

Managed agent subagents are a third lane. They are not a replacement for Conversation AI front desk bots. Use them for scheduled or multi-step CRM jobs in Agent Studio, then keep inbound reception on Conversation AI, prompt or flow.

Still choosing where AI sits in the wider product? The 2026 Explained tour maps menus; Workflow Maps shows how automations connect after agents hand off.

Pros and Cons

FAQs

Verdict

Flow Based Builder V2 is the Conversation AI update to use when inbound bots need ordered steps you can see and fix. Build with AI drafts the graph, questions stop silent wrong assumptions, and the execution timeline makes Test useful. Managed agent subagents are the companion update for scheduled CRM work: one sentence, a planner, and specialists with smaller tool kits.

You do not need to migrate every prompt agent tomorrow. Keep simple bots on prompts. Turn on Labs, build one front desk flow, run Test Mode on a managed follow-up job, and only then decide what to clone with snapshots into client accounts.

Practical rollout for an agency week: day one, enable Labs and rebuild one inbound bot on the canvas. Day two, wire the real calendar and retest the timeline. Day three, create one managed follow-up agent with subagents and fix owner gaps. Day four, package what worked into a snapshot for the next client sub account. That cadence beats rewriting every prompt on Friday night.

Turn this into a system you can sell. Watch the free masterclass for the agency method, or get free access to the community through Greg's GoHighLevel link.

Related guides: Conversation AI Build Agents · Snapshots · Workflow Maps · What Is GoHighLevel · Explained · Get free access

More from the Faceless GHL series lives on YouTube. This page tracks video 13 (Flow Builder V2 + Subagents). HighLevel® is a registered trademark of its owner. This site is not affiliated with or endorsed by HighLevel.