01
Start from a template
Knowledge Base, Datastore Entry, Third-Party Integration, Post-Conversation Analyst — or a blank helper.
How it works
Specialise the hands, keep the brain clear
- 02
Attach the tools
Give it only what the job needs: one knowledge base, a datastore, or the connected apps you pick.
- 03
Test in the playground
Run it, see every tool call, then pin the exact app actions it should use.
- 04
Connect it to agents
Add it to one agent or several. The agent calls it when the conversation needs that job done.
Capabilities
What sub-agents are good at
Sub-agents are how your agent gets tools. Anything that should be accurate, repeatable and separately tunable belongs in one.
Knowledge search
Search one knowledge base and hand back an answer grounded in what it found.
Datastore writes
Save leads, tickets and feedback as records with the exact fields you defined.
Connected app actions
Look up a Shopify order, book on Google Calendar, update HubSpot — using the apps you connected.
Post-conversation analyst
Runs after the chat ends, fills the outcome fields you define, and can use its tools for follow-up.
Its own model
Pick a model per helper, and optionally bring your own OpenRouter or Vercel AI Gateway key.
No-code setup
Instructions, tools and output fields — all configured in the builder, no framework.
In practice
Why teams split work this way
One mega-agent that can do everything usually does nothing reliably.
Clearer conversations
The visitor-facing agent stays focused on talking; it hands a short request to the helper and relays the answer.
Tighter permissions
A helper can only use the tools you attached. Only the helper you gave Gmail can touch Gmail.
Every run is logged
Each run is recorded on the conversation: the request, tools used, the answer and token usage.
Shared building blocks
One well-built helper can be connected to every agent that needs it.
FAQ
Questions
Do I need code?
No. You describe the job in plain English, pick the tools, and the sub-agent runs when called.
Where do sub-agents run mid-conversation?
In website chat and browser voice, your agent can call any sub-agent connected to it. Each run gets up to 5 tool steps to finish the job.
Can one agent use many sub-agents?
Yes. Connect the helpers that match the job — support might get knowledge + tickets; sales might get leads + calendar.
Can my team run them manually?
Yes. While handling a live chat, voice session or Telegram thread, a teammate can pick one of the agent's sub-agents, type a request and run it. The answer is shown only to them, ready to copy — the visitor never sees it.
Can it run after the chat ends?
Yes. Choose a sub-agent as an agent's post-conversation analyst. It reads the transcript and writes a structured outcome with the fields you define, without blocking the visitor. If it returns nothing usable, the default analyst fills in.
What happens when I hit a plan limit?
Knowledge searches, datastore writes and integration calls each have a plan allowance. Once one runs out, that tool is switched off until you upgrade, so the helper never tries a call it can't make.
Free plan · paid plans free for 7 days
Build your first customer agent.
No credit card. Questions? Write to support@embage.com.