> ## Documentation Index
> Fetch the complete documentation index at: https://docs.conduktor.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Conduktor Agents

> Run AI agents inside Conduktor Console that watch your Kafka platform on a schedule or in response to changes, and hand back reports, recommendations and approvals.

Conduktor Agents are AI teammates that run inside Conduktor Console. They help every team keep their Kafka clusters secure, reliable, well-governed and cost-efficient. An agent runs a task against your Kafka platform, on a schedule or when something changes, investigates what it finds and brings you back a finding with its evidence and a next step: a report to read, a recommendation to act on or a fix to approve.

Agents aren't a separate product. They use the same tools as the [Conduktor MCP server](https://docs.conduktor.io/guide/conduktor-in-production/automate/mcp.md), so they work from what Console already knows about your clusters, topics, schemas, connectors, access control and [Insights](https://docs.conduktor.io/guide/insights.md), with Console's permissions applied. You bring your own model, on your own account.

> [!NOTE]
>
> Agents are available from Console 1.48.0 and are turned off by default. In this release, only Console administrators can see and use them. Access for other users is coming later in November.

## How agents work

Four things make up the feature. You'll meet them in this order in the Console UI.

| Concept | What it is |
| - | - |
| **Identity** | Who the agent acts as. Every agent gets its own identity: a machine token with a name, an expiry and its own permissions, linked to a named Console user and limited to that user's permissions. Pick the task and cluster when you create it and Console selects only the permissions that task needs. The task only sees what its identity can see, data is masked as it is for that user, every action is attributed to the identity in the audit log, and you can revoke one identity without touching the others. |
| **Task** | What the agent does. A task is a prompt, a tool allowlist and an output type. Console ships with a catalogue of ready-made tasks, such as cluster summaries, certificate expiry watches and schema drift watches, and administrators can write their own. The **Tasks** page describes each one. |
| **Agent** | A task, an identity and a condition, saved with a name. The condition is either a schedule (hourly, daily, weekly or monthly) or an event: a change to a topic, subject, consumer group or connector. An agent also names the model that runs it. |
| **Output** | What the agent hands back. You decide how far each agent goes. With a **report**, the agent finds and you read. With a **recommendation**, the agent proposes and you act or dismiss. With an **approval**, the agent proposes the exact steps and waits for you to approve or reject. Every output carries a trace of what the agent did to produce it. |

An agent run looks like this:

1. The schedule fires, or a matching change lands in the audit log.
2. Console starts a run as the agent's identity and sends the task to the model.
3. The model calls Console tools to gather what it needs, within the identity's permissions.
4. Console stores the result as an output and shows it in the **Output** section.

Agents never change your Kafka resources on their own in this release. Tasks that propose a change produce an approval for you to decide, and Console records your decision without executing it. Automation, where an agent carries out steps you approved in advance, is coming later in November.

Every run stays within limits, so a looping or runaway agent stops on its own: 20 model calls, 50 tool calls, 500k tokens and 10 minutes per run, and a run that keeps repeating the same call is stopped.

> [!WARNING]
>
> The results of an agent run are produced by the AI model you connect, and are open to interpretation by that model. Review them before you act on them.

## Get started

> [card] [Set up agents](https://docs.conduktor.io/guide/conduktor-in-production/automate/agents/configure.md)
>
> Enable the feature and connect Console to a model provider.

> [card] [Create and run your first agent](https://docs.conduktor.io/guide/conduktor-in-production/automate/agents/get-started.md)
>
> Create an identity, create an agent from a ready-made task and read its first output.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.