Agent Automations
An Agent Automation is like setting up an assembly line where each station is staffed by one of your agents. The work moves down the line on its own — one agent researches, the next writes, the next sends — and the whole line runs again every time it's set off, with no one standing over it. Agents become steps in automated workflows that run without you watching, every time a schedule fires, an outside alert arrives, or an event matches.
Unlike Agent Tasks — which run once on a single record (one row) — Agent Automations are built to keep running on every trigger (the thing that kicks the routine off), often coordinating several agents across multiple steps.
Before you start
- At least one saved agent. If you don't have a specialist for a step, you can use the default Nirvai agent (see below).
- A sense of what should set the routine off — a schedule, an incoming alert from another service, an event, or a button you press. Picking that is covered in Choosing a Trigger.
- You'll build the routine either with AI Setup (you describe it, the AI assembles it) or the Manual Builder (you place each step yourself).
How agents work in an automation
Each step in the routine is handed to an agent. The agent receives whatever the previous step produced (or, for the first step, the data that came in with the trigger), works on it using its own instructions, and passes its result to the next step.
Each agent:
- Receives the previous step's output as context
- Uses its own tools, knowledge, and instructions to do the work
- Produces an output that flows to the next step
Step-by-step: assigning an agent to a step
When building an automation (with either AI Setup or the Manual Builder):
- Add an agent step to the routine.
- Select which agent to assign — you'll see your full agent list.
- Write step instructions that tell the agent what to do with the incoming data.
Step instructions vs. agent instructions
Your agent already has its own instructions that define how it generally behaves. Step instructions are extra notes, specific to this one step:
| Agent instructions | Step instructions |
|---|---|
| "You are a research assistant. Use web search to find information. Present findings in bullet points." | "Search for the latest news about the company mentioned in the input. Summarize the top 3 stories." |
| Always active, across every conversation | Active only during this automation step |
| Define personality, tone, and general behavior | Define the specific job for this step |
Both apply at the same time — the agent follows its general instructions and the step-specific ones together.
Confirming agents (AI Setup)
When you use AI Setup to build a routine, there's an agent confirmation step before it goes live:
- The AI proposes which agent to use for each step.
- You see a card for each proposed agent showing its name, tools, and knowledge.
- You confirm the selection or ask for changes.
- Once confirmed, the AI assembles the routine with those agents.
This way you know exactly which agents will run before the automation goes live.
Run modes
The AI automatically picks a run mode based on how involved your automation is. The mode changes how much depth each agent puts into a step:
| Mode | Agent behavior | Best for |
|---|---|---|
| Agile | Faster, standard depth of analysis | Simple jobs, time-sensitive routines |
| Intensive | Deeper analysis, more thorough responses | Heavy research, detailed reports |
Intensive mode is available on PRO plans. It gives agents more time per step for work that needs deeper analysis.
The default agent
If you don't have a specialist agent for a step, you can use the default Nirvai agent. It handles straightforward jobs well out of the box. For better results on specialized work, use a custom agent with the right tools, knowledge, and instructions.
Multi-agent patterns
Sequential pipeline
Agents work one after another, each building on the previous step's output:
| Step | Agent | Task |
|---|---|---|
| 1 | Research Agent | Find information |
| 2 | Writer Agent | Draft content from the research |
| 3 | Editor Agent | Review and polish the draft |
| 4 | Notification Agent | Send the final content |
Router-based branching
A router step sends the flow to different agents depending on a condition:
Specialist agents
Different agents handle different kinds of work in the same routine:
| Agent | Specialization | Tools |
|---|---|---|
| Research Agent | Web search, data gathering | Web search |
| Data Agent | Database operations | Database tools |
| Communication Agent | Sending messages | Slack, email tools |
Step consolidation
If two steps in a row use the same agent, Nirvai merges them into a single step. This is more efficient — the agent does both jobs in one run instead of two.
Before merging:
| Step | Agent | Instructions |
|---|---|---|
| 1 | Research Agent | "Find the latest news" |
| 2 | Research Agent | "Summarize the findings" |
After merging:
| Step | Agent | Instructions |
|---|---|---|
| 1 | Research Agent | "Find the latest news and summarize the findings" |
If you want two steps to stay separate (for routing or record-keeping), assign a different agent to each step.
Plan limits
| Plan | Max agents per automation |
|---|---|
| Basic | 3 |
| PRO | 10 |
When to use an automation instead of a task
| You want to… | Use |
|---|---|
| Run the same job every morning / whenever an outside alert arrives | Agent Automation |
| Coordinate several agents across multiple steps | Agent Automation |
| Send the flow down different paths based on a condition | Agent Automation |
| Run a one-off job on a specific database row | Agent Task |
| Use a row's own data as context | Agent Task |
See the Tasks & Automations overview for a fuller comparison.
Troubleshooting
| Problem | Fix |
|---|---|
| An agent isn't acting on the right input | Step instructions only cover this step. Make them spell out what to do with the incoming data from the previous step. |
| Two steps merged into one unexpectedly | They use the same agent, so Nirvai consolidated them. Assign a different agent to a step to keep it separate. |
| I can't add more agents to my automation | You've hit your plan's limit — 3 on Basic, 10 on PRO. Reuse an agent across steps, or upgrade. |
| Intensive mode isn't available | Intensive mode is a PRO-plan feature. On other plans, routines run in Agile mode. |
What's next
After this, continue to Practical Examples to see complete automations built with agents. To go deeper on configuring each step, see Agent Steps, and for how automations work overall, see the Automations Overview.