How Do I Build a Multi-Agent AI Workforce with ChatSyncs?

A single, broad AI agent can only be pulled in so many directions before its System Prompt gets messy and its answers get vague. A multi-agent workforce splits the work instead — each agent specialized for one job, with Agent Routing sending each message to the one built for it.

Three pillars, one workforce

Every agent you build rests on three pieces. Skip any one of them and the chain breaks.

Pillar	Job
Knowledge Sources	The foundation — the training data an agent answers from.
AI Agent (System Prompt)	The brain — the exact instructions, what to collect, and which actions to run.
AI Configuration (Agent Routing)	The router — which agent handles which kind of message.

Knowledge Sources — the foundation

This is where an agent’s training data lives: FAQs, URLs, files, a Google Sheet, an API, or uploaded media. Different agents need different knowledge — a Support agent needs troubleshooting docs, a Sales agent needs pricing, a Media agent needs a product photo library. See How to Train Your AI Chatbot for the full walkthrough of each source type.

Keep each agent’s Knowledge Set to what that one agent actually needs. Mixing training data across agents dilutes accuracy and makes routing less effective.

AI Agent — the brain

Agents tab listing three agents, each with a Knowledge Sources count and Active status

The agent draws on its connected Knowledge Sources, but the System Prompt is what actually runs the show — it defines the agent’s exact role, what it must collect, and how it should behave. A basic prompt like “You are a sales agent, help customers buy things” produces a mediocre agent. A strong one is specific about the role, the required information, the conversation rules, and exactly which System Prompt actions to run and when. The golden rule: write the condition first, the action right below it. For example:

When the customer mentions their budget, save it:
##save_custom_field##: Budget

Once the customer shows clear intent to buy, label them:
##add_label##: Qualified Lead

That conditional pattern — plain instruction, then the ##action## it triggers — is what turns a passive chatbot into a digital worker that actually completes tasks.

Worked example: Support, Sales, and Media agents

Three specialized agents, each with a narrow job:

Agent	Job	Knowledge Source	Key actions
Support Agent	Answer questions, escalate what it can’t resolve	URL training (help center / docs)	`##add_label##`, `##assign_human_member##` / `##assign_human_role##`
Sales Agent	Qualify leads, recommend a plan	Content/Text (pricing, plans)	`##save_custom_field##`, `##add_label##`, `##assign_sequence##`
Media Agent	Send product photos on request	Media upload (images + descriptions)	none — answers from Knowledge Sources only

Build at least two agents before going live with routing — seeing them work side by side is the best way to understand how a message actually gets dispatched.

Step 1: Train each agent’s Knowledge Sources

Sales Agent — Content or FAQ

Paste your pricing plans, feature lists, and qualification criteria straight into the Content/FAQ box — the fastest way to get a sales agent answering accurately. (A connected Google Sheet works too, if your pricing changes often and you want it to stay in sync automatically.)

Support Agent — URL

Point it at your help center or documentation pages. Use the Fetch/Remove Content Configuration selectors to strip navigation, footers, and ads so only the real troubleshooting content gets learned.

Media Agent — Media

Upload product photos directly under Media, with a short, accurate description for each one (e.g. “Red cotton kurti, size M, price 1200”). ChatSyncs uses these descriptions to match the right photo to what a customer asks for.

See How to Train Your AI Chatbot for the full steps and screenshots for each source type.

Step 2: Write each agent’s System Prompt

Create AI Agent form with Name, Slug (auto-generated from name if empty), Status, Description, and System Prompt fields

Follow How to Create Your First AI Agent once per agent. Here’s a realistic Sales Agent prompt, as a starting template:

You are the Senior Sales Agent for [Your Business].

ROLE
Understand the customer's needs, recommend the right plan, and qualify them as a lead.
Be professional, persuasive, and helpful.

REQUIRED INFORMATION
Before recommending a plan, naturally collect:
1. Business type
2. Primary communication channel
3. Main requirement
4. Expected monthly volume
5. Approximate budget

CONVERSATION RULES
- Ask only one or two questions at a time.
- Don't repeat questions already answered.
- Never invent pricing or features not in your training data.

SAVE INFORMATION
When the customer mentions their budget, save it:
##save_custom_field##: Budget

QUALIFICATION & ACTIONS
Once you've collected the required info and the customer shows clear intent to buy:
##add_label##: Qualified Lead

If they're interested but need more time:
##assign_sequence##: Sales Follow-Up

HUMAN HANDOFF
If they ask about enterprise or custom pricing, or explicitly ask for a human:
##add_label##: Sales Escalation
##assign_human_role##: 3

And a Support Agent prompt — simpler, focused entirely on answer-or-escalate:

You are the Customer Support Agent for [Your Business].

ROLE & KNOWLEDGE USAGE
Answer questions about features, policies, and troubleshooting from your connected
Knowledge Source. Be patient, clear, and empathetic.

HUMAN HANDOFF (ESCALATION)
If the customer reports a billing issue, has an unresolved technical problem after
troubleshooting, or explicitly asks for a human, escalate:
##add_label##: Escalated Support
##assign_human_role##: 5

Because the agent collects context before escalating, the human teammate who picks up the conversation already has what they need — no re-asking the customer to explain the issue again.

Step 3: Set up the Media Agent

Only one confirmed way to send images from an agent: upload them under Knowledge Sources → Media, with a description for each. The agent fetches the matching photo when a customer asks for it — this is what ChatSyncs’ own System Prompt Actions reference confirms (“Do not put image URLs in the prompt. Add product/gallery images under Knowledge Sources (Media).”). Some BotSailor-based tutorials describe a second method — typing image URLs directly into the System Prompt with a ##send_image##-style action — but that isn’t confirmed for ChatSyncs and contradicts the in-product guidance, so it’s left out here until confirmed.

You are the Agent for Sending Product Photos for [Your Business].

ROLE
Provide accurate product images. Keep text brief — let the images do the talking.

CONDITIONAL DELIVERY RULE (CRITICAL)
Never send images automatically at the start of a conversation. Hold them in reserve and
deliver them only when a customer explicitly asks to see a photo.

KNOWLEDGE USAGE
When a customer asks for a specific product ("show me the red kurti"), fetch the matching
image and description from your connected Knowledge Source.

Step 4: Activate agents and set up routing

Active Agents multi-select dropdown with three agents all selectable

On AI Configuration, add all three agents to Active Agents, then add one Agent Routing rule per agent:

Message contains	Routed to
`pricing, plan, upgrade, budget`	Sales Agent
`bug, not working, error, refund`	Support Agent
`photo, image, show me, what does it look like`	Media Agent

Routing reads intent, not just exact keywords — “How much would this cost for a small business?” routes to the Sales Agent without containing the word “pricing.” And if a customer’s intent shifts mid-conversation, the conversation can hand off to a different agent’s rule automatically — see Agent Routing for how transfers work.

Step 5: Finish the global settings

A few settings from How to Create Your First AI Agent are worth setting deliberately for a multi-agent setup:

Contextual Memory — on, so agents remember earlier messages instead of asking the same question twice.
Enable Typing on Indicator — on, for a more natural pace between message and reply.
Restricted Topics — set per agent so each one stays inside its lane — see How Do I Restrict AI on Specific Keywords?.

Click Save Settings. All active agents now run in parallel, and every new conversation is routed the moment it arrives — no manual sorting.

Test before going live

Run each agent through a full conversation before activating it for real customers. Confirm it:

Answers accurately from its own Knowledge Source, not another agent’s.
Asks for missing information naturally, without repeating questions already answered.
Saves data to the correct custom fields.
Triggers the right action at the right condition — not too early, not missed.
Sends images only when explicitly asked (Media agent).
Escalates or labels a contact without duplicating the action on a retry.
Handles an ambiguous or unexpected message without inventing an answer.

Why this is worth the extra setup

Precision training eliminates noise — each agent only draws from its own Knowledge Source, so the Support agent never confuses a pricing question with a refund policy.
Autonomous actions replace manual work — labeling, escalation, and follow-up enrollment happen the moment their condition is met, without a human reviewing every conversation first.
Conditional media delivery keeps conversations clean — images only arrive once a customer actually asks, instead of cluttering every reply.

Frequently asked

How many agents should I actually build?

Start with one broad agent. Split into specialized agents once you notice one prompt trying to juggle clearly different jobs — e.g. sales vs. support vs. sending media — each with different knowledge and different actions.

Can I send an image directly from the System Prompt with a URL?

Not confirmed for ChatSyncs — the product’s own System Prompt Actions reference says not to. Upload images under Knowledge Sources → Media instead; the agent fetches them from there.

What stops two rules from both matching the same message?

Write routing rules around distinct message types so overlap is rare. Where a message could reasonably match more than one, put the more specific rule first — see Agent Routing.

Can one agent hand off to another mid-conversation?

Yes — either through matching a different Agent Routing rule as the conversation moves on, or explicitly via a ##transfer_agent## action in the System Prompt. See AI Agent Actions.

Do I need to build all agents before testing any of them?

No, but build at least two before testing routing specifically — with only one agent active, every message goes to it regardless of any rule, so you can’t actually see routing decide anything.

​Three pillars, one workforce

​Knowledge Sources — the foundation

​AI Agent — the brain

​Worked example: Support, Sales, and Media agents

​Step 1: Train each agent’s Knowledge Sources

​Step 2: Write each agent’s System Prompt

​Step 3: Set up the Media Agent

​Step 4: Activate agents and set up routing

​Step 5: Finish the global settings

​Test before going live

​Why this is worth the extra setup

​Frequently asked

Three pillars, one workforce

Knowledge Sources — the foundation

AI Agent — the brain

Worked example: Support, Sales, and Media agents

Step 1: Train each agent’s Knowledge Sources

Step 2: Write each agent’s System Prompt

Step 3: Set up the Media Agent

Step 4: Activate agents and set up routing

Step 5: Finish the global settings

Test before going live

Why this is worth the extra setup

Frequently asked