Heymarket AI Agents Heymarket AI Agents can draft or send replies to incoming messages, depending on whether you use Supervised Mode or Autonomous Mode, drawing on a knowledge base you control and instructions you write. They get better with context and iteration, so the goal of your first launch is not perfection. It is a working agent in front of real conversations, in a mode where humans still verify every message send.
This article walks you through a minimum viable setup. Plan for about 30 minutes.
What an AI Agent does
An AI Agent is a configurable assistant assigned to one or more of your Heymarket inboxes. When a contact sends a message, the agent reads the conversation, consults your knowledge base, and either drafts a reply for a teammate to review (Supervised Mode) or sends it automatically (Autonomous Mode). Admins control the agent's identity, tone, knowledge sources, response length, and rules for handing off to a human.
How the agent gets triggered
The AI Agent responds to inbound messages in any conversation where it is the assignee. There are two ways to make it the assignee:
- Auto-assign new conversations. Turn on "Assign all new conversations to AI Agents" at the inbox level so every incoming chat goes to the agent automatically.
- Manually assign a conversation to the AI Agent. Open any chat, switch the assignee to AI Agent, and the agent will reply to the next inbound message.
You do not need auto-assignment to start using the agent. Manual assignment is a great way to test it on a single chat or run a soft launch. Either way, the trigger is the same: a contact sends a message, and the agent (as the assignee) drafts or sends a reply. If a teammate replies but leaves the AI Agent assigned, the next inbound contact message can still trigger the agent.
Outbound messages from a human teammate do not trigger a follow-up reply from the agent.
Send a test message and see the agent respond
The quickest way to see the agent work end to end takes about two minutes:
- Open an inbox where the AI Agent is a member.
- Open an existing chat or start a new one with your own phone number.
- In the assignee dropdown at the top of the chat, select AI Agent.
- From the contact side (your phone, or another test number), send an inbound message such as "What plans do you offer?"
- Watch the agent generate a reply. In Supervised Mode it appears as a draft for review. In Autonomous Mode it sends automatically.
If nothing happens, check that the agent is enabled, the agent has at least one knowledge source, the agent is a member of the inbox, the chat is assigned to the AI Agent, and the latest message is inbound.
The eight-step lifecycle
- Decide what the agent is for (Article 2)
- Write its instructions (Article 2)
- Build its knowledge base (Article 3)
- Configure its behavior (Article 4)
- Assign it to an inbox in Supervised Mode (Article 5)
- Test before going live (Article 6)
- Review suggestions and feedback (Article 7)
- Iterate, then graduate to Autonomous Mode when ready (Article 7)
Minimum viable setup
For a first launch, keep the scope tight:
- One primary goal (for example, answer common support questions)
- One knowledge source (your help center URL or a single PDF)
- A short instruction set covering identity, objective, tone, and constraints
- The two default unassignment rules (asks for a human, expresses frustration)
- Supervised Mode
- One inbox
You can add agents, sources, rules, and inboxes later. Starting narrow makes problems easier to spot and fix.
What success looks like in your first week
- Suggested replies are displayed shortly after an incoming message
- Most suggestions are accurate enough to send with light edits
- Unassignment rules fire correctly when a contact asks for a human
- The Agent Enhancements page surfaces a small queue of knowledge gaps to address
Next steps
Continue to Article 2: Define Goals and Write Effective Instructions.