---
title: "Choose an agent mode"
description: "Choose Plan, Research, or Eval for your study, then review the plan or findings before continuing the work."
---

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

# Choose an agent mode

Choose an agent mode to plan a research task, carry it out, or check an existing result. The mode sets the agent's role and working instructions.

The model is the AI system that produces the response. It has a separate selector, such as **GPT Terra** in these screenshots.

Before you begin, open your study project and a conversation. Follow [Start a session](/projects-and-files/start-a-new-session/) for a separate task. The examples use a fictional study for a national grocery retailer.

For the examples, [send the practice request](/projects-and-files/start-a-new-session/#send-your-first-request) using `study-brief.txt`. Use the same conversation when you plan revisions or evaluate the questions you received.

## Select a mode in the app

1. In your InsightGig conversation, find the current mode beside **Ask anything...**. The message box and its task controls form the composer.

2. Select the current mode to open the list. A check mark identifies the current choice. Select **Research**, **Plan**, or **Eval**.

   ![Mode picker showing Research selected and offering Plan and Eval](@/assets/images/work-with-agents/agent-modes-current.png)

3. Check the mode name in the composer. Then check the model and thinking level before sending your request.

   ![Composer showing Plan selected beside the model and thinking controls](@/assets/images/work-with-agents/agent-mode-plan-current.png)

Your next request uses the selected mode. You can switch again in the same conversation when the next task needs a different role.

## Match the mode to your task

| Mode | Start here when | Main result |
| --- | --- | --- |
| Plan | The task needs decisions about sources, methods, audience, or outputs. | A plan for you to review before execution. |
| Research | You want the agent to investigate, analyze, or produce an output. | An answer, analysis, or requested deliverable. |
| Eval | You have an output and want it checked against sources and requirements. | A review report with findings and recommended fixes. |

### Use Plan to define the work

Use **Plan** when a request such as "analyze this study" leaves important choices open. It reads relevant sources before asking questions that affect the result.

Those questions can cover the audience, research objective, included files, output format, and quality checks. Answer them with the decisions you want reflected in the work.

Plan saves execution instructions in `PLAN.md`, a text document in the conversation's planning area. It can save inspection results there too. Its role is to prepare the work without changing your study sources or producing the final deliverable.

For example, ask for a plan to revise the questions for a grocery checkout study.

![Draft request in Plan mode asking for a revision plan before the questions are rewritten](@/assets/images/work-with-agents/agent-mode-plan-draft.png)

Open the plan link in the reply to check the sources, the intended output, and the review criteria.

![Saved PLAN.md showing the objective and requirements for the retail interview questions](@/assets/images/work-with-agents/agent-mode-plan-report.png)

Review the proposed sources, method, outputs, and checks before approving execution. Follow [Plan and run a task](/work-with-agents/agent-controls/plan-and-run-a-task/) for the planning walkthrough.

### Use Research to carry out the task

Use **Research** for a defined research request or to execute an approved plan. It investigates sources and can use specialist agents for numerical analysis or document synthesis.

Research can create or change project files when the task calls for it. Say whether you want an answer in the conversation or a saved file, and name the source material.

For example, draft interview questions from a study brief.

```txt
Read study-brief.txt and draft three neutral interview questions.
Focus on shoppers who abandoned their first online grocery order.
Name the issue each question explores and reply in the conversation.
```

Research can pause for decisions while working, such as approving an outline or written content before creating a report. Answer the request shown in the conversation.

When the task needs external sources, follow [Choose Web or Desk Research](/work-with-agents/agent-controls/choose-web-or-desk-research/) for the conversation controls.

### Use Eval to check an output

Use **Eval** after an answer, report, presentation, or workbook exists. Name the output, provide its sources, and say what needs checking.

It can compare claims with source evidence, check calculations and quotations, and assess coverage of the brief. It can also check whether an output is usable for its intended audience.

Evaluation saves its findings in `EVALUATION.md`, a review report in the conversation's evaluation area. It can create verification files there, while leaving study sources and the reviewed deliverable unchanged.

For example, ask Eval to check the interview questions against the brief for leading wording and missing topics.

![Draft Eval request asking for a review of three interview questions against the study brief](@/assets/images/work-with-agents/agent-mode-eval-draft.png)

Open the review link in the reply to see what Eval checked. The screenshot shows two questions that assumed what happened during checkout. Your findings may differ. Compare each flagged question with the brief before deciding what to revise.

![EVALUATION.md open to the verdict and findings for the retail interview questions](@/assets/images/work-with-agents/agent-mode-eval-report.png)

## Follow handoffs between modes

A handoff lets another mode continue the same conversation. You can select the next mode yourself or respond if the agent offers to switch.

1. Review the plan. When you approve it, select **Research** and ask it to carry out the plan, or accept the switch if Plan offers one.
2. Open **Eval** after Research produces the output. Ask it to check the output against the sources and your review criteria.
3. Read the findings and decide which fixes to make. Return to **Research** and name the fixes to apply.

For a defined task, you can start directly in **Research**. Review claims and calculations against your sources before using the output.

## Next steps

Follow [Choose a model and thinking level](/work-with-agents/agent-controls/choose-model-and-thinking-level/) for model settings. Read [Review requests for approval](/work-with-agents/agent-controls/choose-approval-option/) to control permission requests during the task.

Source: https://docs.insightgig.ai/work-with-agents/agent-controls/choose-agent-mode/index.mdx
