Act Overview
Act is the agentic AI layer of the HiveMQ Platform.
In Act, you build and deploy AI agents that act on your real-time IoT data, without application code.
Open Act from the top navigation bar.
| Act is currently available as a public preview. |
Agents connect directly to your MQTT brokers, databases, and APIs, analyze the data they observe, and take automated actions.
You describe what an agent does, and the platform runs it, keeps it alive, and manages your fleet (all agents that you deploy across your networks). When an action needs approval, the platform routes it to a person.
What You Can Build
With Act, you can build agents for tasks such as the following:
-
Monitor industrial sensors: Subscribe to MQTT topics, detect anomalies, and publish alerts.
-
Automate threshold responses: Evaluate conditions with rules or a large language model (LLM), and act when limits are exceeded.
-
Add human review to quality control: Query a database for defect data, then route critical findings for human review before the agent acts.
-
Connect to your enterprise resource planning (ERP) or supervisory control and data acquisition (SCADA) system: Call REST APIs, write results back, and create maintenance tickets.
To build your first agent end-to-end, see Get Started.
How an Agent Works
Every agent runs the same four-stage loop on each cycle: Sense → Reason → Actuate → Reflect.
| Stage | What Happens |
|---|---|
Sense |
Collect data from MQTT topics, databases, and APIs. |
Reason |
Analyze the data with an LLM or deterministic rules, plan actions, and apply governance checks. |
Actuate |
Execute the planned actions: publish MQTT messages, write to a database, call an API, or send email. |
Reflect |
Learn from the outcomes of a cycle, save memory for the next cycle, and capture metrics. |
An agent runs on a trigger such as an interval, a cron schedule, or an incoming MQTT message and loops continuously until you stop it.
Agents Are Configuration, Not Code
An agent consists entirely of configuration, called a data-defined agent (DDA). You do not write code or run a build step. The platform reads the configuration and runs the agent for you.
An agent can reason in two ways. You can use one mode or both in the same agent:
| Mode | When to use |
|---|---|
LLM-powered |
Pattern analysis, anomaly detection, and natural-language reasoning. |
Deterministic |
Fixed threshold rules, with no AI cost or latency. |
Use deterministic rules for straightforward checks. Call an LLM only for an analysis that needs it.
Networks and Orchestrators
Networks organize agents, and orchestrators run them:
-
A network groups everything for one site, line, or environment, for example
berlin-factoryormunich-staging. Tools and resources, such as databases, APIs, and Model Context Protocol (MCP) servers, that you register on a network are available to every agent on the network. You configure a shared connection once rather than per agent. See Create a Network. -
An orchestrator is the software you run on your own infrastructure to host agents alongside your brokers. Every orchestrator belongs to one network and runs your agents as Docker containers. See Deploy an Orchestrator.
The Control Plane is the agent-management layer. Use the Control Plane to create networks, store agent templates, deploy agents, and monitor agents.
The Control Plane is part of the HiveMQ Platform, not a separate product.
To access the Control Plane, open the Act tab.
Two Brokers: Your Data and the Agent Bus
Each agent maintains two separate connections. This separation keeps your business traffic apart from coordination traffic:
| Connection | What it Carries | Configured By |
|---|---|---|
Data connections |
Business data (sensor readings, database results, API responses) |
Your agent configuration |
Agent bus |
Inter-agent messages and human feedback requests |
The orchestrator, automatically |
Because the two connections are independent, a problem with your data broker does not disrupt agent-to-agent communication, and vice versa.
How Deployments Reach Your Infrastructure
You never push files to your servers. Instead, the orchestrator periodically checks in with the Control Plane to ask (or pull) what it should run. The orchestrator then starts or stops containers to match the list.
The pull model has two effects:
-
All connections flow outbound from your infrastructure. Orchestrators and agents never open listening ports, so they work behind firewalls and Network Address Translation (NAT).
-
You deploy, update, and remove agents from the Act tab. The orchestrator applies each change on its next check-in.
For the full deploy flow and agent lifecycle states, see Deploy an Agent. To watch deployed agents run, see Monitor Agent Health.
Where Act Fits
Each platform layer builds on the previous one:
-
Connect provides the brokers that carry your MQTT traffic.
-
Contextualize defines the structure that the traffic must follow.
-
Analyze verifies that the traffic follows the structure.
-
Act runs AI agents on the verified data.
Act consumes the governed data streams that the other layers prepare. Before you point an agent at a data stream, check the health of the stream in Analyze.
| Connect, Contextualize, and Analyze are not prerequisites for Act. The other layers give your agents governed, well-structured data, so HiveMQ recommends them, but you can use Act on its own. For example, you can run agents without a broker connected to the platform. |