# What MOI agents can do

An agent combines a model, working instructions, and available resources into a reusable AI assistant. Give it your goal and materials; it can read files, retrieve knowledge, or call tools as needed, then return results in the conversation.

## Start with a real task

For example, upload a contract and a supplementary email to a new conversation. Ask the agent to organize their commitments into a table with four columns: item, agreement, source, and questions to confirm. Review each row against the attachments. Continue the conversation to add omissions, explain evidence, or change the format.

## Common uses

| Goal | Starting point | When to add configuration |
| --- | --- | --- |
| Summarize, compare, or organize files | Upload attachments in a new conversation | Create a knowledge base when the material must be reused over time |
| Answer questions about policies or manuals | Use an agent with a bound knowledge base | Add reference material and business definitions |
| Follow a repeatable analysis or writing method | Configure the system prompt and bind skills | Create a skill when several agents need the same method |
| Search public information or access business systems | Configure and bind the required tools | Ensure credentials, network access, and business permissions are available |
| Execute code or process files | Select a runtime environment in the conversation | Use a specific environment for compute or network requirements |
| Produce daily summaries or respond to events | Configure automation | First verify the same agent can complete the task in a conversation |

## You do not need to create an agent first

Open **New conversation** and use an agent provided by your workspace. [Create your own agent](../build-in-conversation/create-agent-by-chat.md) when you need lasting responsibilities, output formats, or resource boundaries.

Available capabilities depend on the model, configuration, resource permissions, and runtime environment. Check the answer, cited materials, and actual execution results before considering a task complete. A completion message alone is not sufficient evidence.

Next: [Complete your first conversation](../build-in-conversation/start-a-conversation.md).