# Which builder for which agent

Match the agent you need, from one that only answers questions to one that checks other agents, to the Studio builder and the setting that produces it.

Studio has three builders, and between them they produce ten forms of agent. This page maps each form to the builder that makes it and the one setting that decides it, so you can pick before you open **New agent**. Each builder has its own page with the full procedure.

**Who can use it:** the **Builder** role or higher, in a tenant licensed for agents. The controls in the last row belong to administrators.

## How it works

Two things separate one form from another: what starts the agent, and who decides its next step. A person's message starts a conversational agent. A clock, a call from another system or a posted event starts a workflow. The order of the work is fixed by you in a workflow, and chosen by an agent when one agent calls or routes to others.

| The agent you need | Builder | What decides it |
| --- | --- | --- |
| Answering agent: answers from your documents and changes nothing | **Answer & assist**, **Search your docs** | The knowledge sources you attach in **Knowledge** |
| Acting agent: answers and also carries out the request | **Answer & assist**, **Answer questions** | The tools you pick in **Tools** |
| Scheduled agent: runs at set times | **Automate a process** | A schedule trigger with a **Cron expression** |
| Event-triggered agent: starts when another system reports something | **Automate a process** | A webhook trigger with a **Path** |
| API agent: another application calls it and gets a result back | Any builder | **REST API** under **Where it runs** when you publish |
| Autonomous agent: given a question, it plans its own steps | **Answer & assist**, **Research report** | The **Research question**, the sources and the depth |
| Agent pipeline: stages in an order you fix | **Automate a process** | The steps, branches and approvals of the flow |
| Manager agent: one agent calls others inside its turn | **Answer & assist**, **Answer questions** | The agents you attach in **Tools** |
| Agent team: each message goes to the right specialist | **Coordinate a team** | The **Specialists** you attach and the **Router prompt** |
| Guardian: checks what agents take in, give out and pass to tools | The Control Room | A guardrail profile, tool-call policies and **Approvals** |

### Conversational agents

Both conversational forms come from the **Answer & assist** wizard. An answering agent reads the knowledge sources you switch on and nothing else. An acting agent adds tools from the MCP servers registered in the Control Room, and write actions get an approval gate by default. See [Build an answer agent](/agents/build-an-answer-agent).

### Workflow agents

A workflow has a single trigger: manual, a schedule or a webhook. A schedule takes a five-field **Cron expression** in the workflow's timezone. A webhook takes a **Path**, and a published workflow gets a trigger endpoint that starts a run when a message is posted to it. See [Automate a process](/agents/automate-a-process).

Any agent becomes an API agent when you publish it with **REST API** ticked under **Where it runs**. The **Deploy** tab shows the address callers use with a bearer token. See [Test, publish and versions](/agents/test-publish-and-versions).

### Autonomous agents

**Research report** builds an agent that turns a brief into a plan, searches the sources you chose and writes a cited report. A run is asynchronous and takes minutes. See [Research report](/agents/research-report).

### Several agents working together

There are three ways to put agents together, and they differ in who decides the order.

- You decide, in advance: a workflow whose steps run in the order you set, with **Add approval** where a person must agree. A step can be handed to an agent from the canvas palette.
- One agent decides, inside its turn: an answer agent with other agents attached in **Tools**. An attached agent answers inside the turn.
- A coordinator decides, per message: **Coordinate a team** routes each incoming message to one specialist. See [Coordinate a team](/agents/coordinate-a-team).

### Checking other agents

The platform checks agents with three controls instead of a separate agent. A guardrail profile runs rails on what goes into an agent, what comes out and what it passes to tools. A tool-call policy either blocks a tool or parks each call until a person approves it. Parked calls wait in **Approvals**. See Guardrails, Policies and Approvals.

## Limits and behaviour

- Only a workflow can run on a schedule. For a research agent, **On demand** is the available trigger.
- A research agent's time and cost budgets are targets saved with the agent. They do not stop a run.
- A coordinator hands the whole turn to one specialist. Specialists do not pass work to each other, and a workflow cannot take a delegated turn.
- A coordinator needs at least two specialists. Its **Confidence threshold** and **Fallback** are saved but not enforced at run time.
- In an agent's **Guardrails** section, only the selected profile's rails and their actions are enforced at run time.
- To have one agent review another's finished work, add a checking step to a workflow and an approval after it.

## Related

- [Create an agent](/agents/create-an-agent)
- [Build an answer agent](/agents/build-an-answer-agent)
- [Automate a process](/agents/automate-a-process)
- [Coordinate a team](/agents/coordinate-a-team)

Source: https://docs.katonic.ai/agents/choose-a-builder
