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

# Choose an approach

> Use the smallest execution primitive that fits your application.

Start with `Agent` for a model that can answer questions and call tools. Add another primitive when you can name the requirement it serves. They can be combined: a workflow can contain agents, and a harness can drive an Agent.

| Requirement | Start with | Why |
| - | - | - |
| Answer a request and choose tools as needed | [Agent](/agents/overview) | The model decides which tool to call within configured limits |
| Package reusable abilities with host grants and aggregate budgets | [Harness](/harness/overview) | Definitions, drivers, and runtime controls are explicit |
| Follow a known sequence with conditions, loops, or retries | [Workflow](/workflows/overview) | The application defines the execution graph |
| Route to or coordinate multiple specialized agents | [Team](/teams/overview) | Members collaborate through a selected coordination mode |
| Deliver work through Redis workers | [Queue](/queue/overview) | Producers and workers separate submission from execution |
| Persist task ownership, approvals, or recovery state | [Durable execution](/durable/overview) | A driver and store implement an explicit replay/reconciliation contract |

## A support assistant, at different stages

**An Agent is enough** to answer a shipping question, search policies, and call your order API. The tutorial begins there.

**A Workflow helps** when a return must follow fixed steps: validate the order, check eligibility, obtain approval, then invoke the payment service. The model may help classify the request; application code enforces the sequence.

**A Team helps** when specialist behavior is the requirement: route a billing question to a billing agent or coordinate several analyses. Each member still needs scoped tools and a clear stopping condition.

**A Harness helps** when several deployments share a support ability but have different grants, runtime drivers, context, or budgets. Its canonical history and the Agent's backing storage have separate ownership.

**Durable execution helps** when an approval or external operation may outlive the process. A queue alone can redeliver a job; the durable driver must decide whether to resume, skip, or reconcile an operation that may already have happened.

## Three decisions to make first

* **Who chooses the next step?** A model-led Agent, a host-defined Workflow, or a Team coordinator.
* **How long does the work live?** One connection, a background job, or a persisted task across restarts.
* **Who may do what?** Bind tools and resources to verified identity and host policy.

If the answers are one model, one request, and a few tools, follow the [quickstart](/quickstart). You can add structure when the application needs it.


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