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

> ## Agent Instructions
> Choose User guide for browser tasks, Developers for CLI, SDK and external agents, and Administration for organization management.
> Confirm identity and organization/project context before operations. Documentation access does not grant ThinkPlatform permissions.
> Review side effects, verify changes with a read, and do not blindly retry an operation with an uncertain outcome.
> Use documented UI steps for the web app and documented public CLI or SDK interfaces for automation. Follow each page’s prerequisites and authentication requirements.

# Create and configure an agent

> Set up an agent and test it with a small request.

An agent combines instructions with a model and the tools, skills, and knowledge it can use. Start with a narrow job, then add access as needed.

## Before you begin

Select the project where the agent belongs. You need access to create agents and to any resources you attach.

## Create and test

1. Open **AgentOps → Agents** and select **Create agent**.
2. Enter a name and an **Instruction** describing the job, expected output, and boundaries. For example: “Summarize the text I provide in three bullets. Do not change any external resources.”
3. Choose an available **Model** and execution mode. Legacy uses an existing Workspace; Sandbox uses a dedicated Workspace and user permissions. The mode is not an ordinary editable field after creation; an existing Legacy agent can use the migration workflow below. Leave tools and knowledge unselected for this simple test.
4. Select **Create**. Find the new agent in the list and open its chat.
5. Send: “Summarize this: The team shipped the search page Monday. On Tuesday, users reported slow results. The team plans to add caching on Friday.” A plausible response contains three bullets: the search page shipped Monday; users reported slowness Tuesday; caching is planned for Friday. Check the facts and bullet count. See [Chat with an agent](https://docs.thinkplatform.ai/guides/agents/chat-and-results.md).

## Configure capabilities

Open the agent's settings to change its instruction, model, tools, skills, or knowledge. For a Sandbox agent, you can also edit its roles and individual permissions; ordinary editing does not change its execution mode. Use **Edit tools** to add only the tools the job needs, and **Edit skills** to attach reusable instructions. Under **Knowledge**, choose available **Table DB** or **Vector DB** resources when the agent needs them. For Sandbox, select available roles or individual permissions appropriate to the work; choices exceeding your permissions are unavailable. Review each tool's **Ask first** setting: it controls whether confirmation is required before that tool runs. Save your changes and test the specific capability with a limited request.

Bundled skills run through the Shell built-in tool. Outside Sandbox execution, add Shell when using a bundled skill; the settings show a warning if it is missing. See [Project skills](https://docs.thinkplatform.ai/guides/agents/project-skills.md) for uploading and attaching project skills.

If creation or a test fails, read the displayed message and check the selected project, access, and configuration. If a test could have changed a resource, inspect its destination before retrying. Review any requested decision as described in [Approvals and recovery](https://docs.thinkplatform.ai/guides/tasks/approvals-and-recovery.md).

## Migrate an existing Legacy agent

Use **Migrate Agent** in the agent's settings when Sandbox is available for the project. This workflow retains the Agent ID and integrations while moving execution to its own Workspace.

1. Save any pending settings edits, then choose **Migrate Agent** to preview the move.
2. Review the selected roles, individual permissions and displayed environment-variable names. If a runtime-reset acknowledgement appears, read it before confirming: local files and manually installed packages remain in the old Workspace and are not copied automatically by this migration. Preserve needed output in durable storage first.
3. Choose **Migrate** and inspect completion. If interrupted, use **Check status**, then the displayed **Complete migration** or **Cancel migration** action as appropriate. Do not start another migration without checking the existing state.
4. After completion, send a small read-only request and verify the agent's expected capabilities. Inspect any denied operation against its new roles and resources.

## Update a Sandbox

An eligible agent can show **A Sandbox update is available** in its settings. Choose **Update** only after reviewing active work and saving required outputs. The workflow can wait for active tasks, prepare a new Sandbox and move workspace files. Manually installed software needs to be set up again. Check for **Sandbox updated**, then verify a small task. If it reports a failure, inspect its status; the interface provides **Retry** and says when the current Sandbox can continue to be used.

A Sandbox update and a Legacy-to-Sandbox migration have different file-transfer behavior. Neither replaces a durable backup. See [Data management](https://docs.thinkplatform.ai/admin/data-management.md).


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