# Plan and start research

Define the question and limits before starting autonomous work. With **New Project**, you approve the project plan before choosing run settings with **Start Marvin**.

## Define the scope

State:

- the question you want answered
- the goals or decisions the work should support
- specific hypotheses that should be tested
- constraints, exclusions, or required methods
- papers, datasets, or prior work that should be used

Review the initial plan against those inputs. Use **Refine with Marvin** to request a change, then read the revised plan. Select **Approve and launch** only when the plan reflects the intended work. Confirming **Launch project?** opens the project; it does not start a run.

You can edit the project name, mission, and goals later. Change the scope before starting another run if the existing goals no longer describe the work.

## Choose a run type

Open the project and select **Start Marvin**.

| Run type | Choose it when |
| --- | --- |
| **Literature** | The main task is to search, review, and synthesize sources. |
| **Analysis** | You want Marvin to analyze data you supplied. |
| **Full** | The project requires literature work, hypotheses, analysis, and experiments together. |

Some run types depend on the project owner's subscription plan. The modal shows the options available for the project.

## Choose thoroughness

| Setting | Behavior |
| --- | --- |
| **Focused** | Stays on the primary direction and folds in work that strengthens it. |
| **Balanced** | Uses judgment to triage lower-likelihood or tangential directions. |
| **Exhaustive** | Explores useful directions more broadly within the project's scope. |

Thoroughness changes how broadly Marvin explores. It does not guarantee that every relevant source or explanation will be found.

## Set the iteration limit

The Free plan runs one iteration at a time and does not show **Run to the cap**. On plans with iteration controls, set **Number of iterations** from `1` to `100`, or select **Run to the cap**. **Run to the cap** submits a maximum of 100 iterations. It is still bounded by time and budget, and Marvin can stop earlier.

Start with one iteration when you want to review the direction before committing more time or spend. You can start another run later.

## Choose compute and files

**Literature** disables experiment compute controls because it performs search and synthesis without experiment compute.

For **Analysis** and **Full** runs, the modal can offer:

- **Max CPU cores**
- **Allowed GPU types**
- **Experiment compute budget**

The **Experiment compute budget** limits experiment compute for this run. It is separate from model usage and account balance. See [Plans, usage, and billing](/docs/plans-and-billing).

Before an **Analysis** or **Full** run, upload needed data under **Data**, then name the file and intended work in the plan or chat. Uploading a file does not queue work. See [Use datasets and project files](/docs/data-and-files).

## Check before starting

**Start Marvin** remains unavailable while Marvin is resolving run status or cannot read the project's settings. It can also be blocked by invalid compute selection, access limits, or account funding.

Before you retry:

1. Read the message in the start dialog.
2. Confirm the run is not already active in another tab.
3. Check your plan and available balance if the message concerns access or funding.
4. Correct the setting, then select **Start Marvin** again.

If the dialog returns warnings after accepting the run, read them before closing it. Once accepted, use **Overview** and **Activity** to follow the run. See [Chat with Marvin and guide research](/docs/guide-research).