# Research tasks

Give an agent a bounded job on a copy of your project, then review and apply selected files.

Research tasks let an assistant work on a copy while you keep control of the manuscript. Give it a job, follow its progress, and bring back the files you want.

## Create a task

1. Open **Research workspace → Tasks**.
2. Create a task and choose a starter, or write your own prompt.
3. Give the task a title. Describe the evidence to use and the files it should produce.
4. Select an available **Agent and model**, then choose **Create task**.
5. Start the task when it is ready.

| Starter | A useful brief |
| --- | --- |
| Literature review | Find recent work for one research question and save a source table with identifiers |
| Evidence audit | Check the claims in one section against the supplied references |
| Analysis | Summarize a dataset and save the script, results table, and interpretation |
| Manuscript revision | Revise a named section against specific feedback |
| Response to reviewers | Map each reviewer point to a response and a manuscript change |

Keep the brief small enough to review. Name the relevant source files or [linked folders](/docs/linked-folders/) and say what should remain unchanged.

## Follow progress

The task queue records waiting, running, review, completed, failed, and cancelled work. Open a task to read its session and inspect outputs. You can edit queued work, cancel a run, or retry a task. Dependencies keep one task waiting until the work it needs is ready.

## Review the result

When a task reaches review, inspect its changed files and diffs. Select the files you want to bring into the manuscript, then apply them. Discarding a result leaves the task history available.

Oleafly checks the destination against the version the task started from. If you have since changed a file, it refuses the conflicting apply and keeps the result available for review. Reconcile that file before trying again.

## Choose an available agent

Use a configured built-in model or a supported [CLI agent](/docs/cli-agents/). The picker shows availability for the current platform and runtime.

Isolated CLI tasks use the operating system's isolation tools on macOS and Linux. They are unavailable on Windows; Codex isolated tasks are also unavailable on macOS with the current runtime profile. Interactive assistant sessions are a separate mode.

## Chat or task?

Use [chat](/docs/ai-chat/) for a conversation and edits in the active project. Use a task when you want a separate copy, a recorded job, and a final file review. Helpers delegated during an ordinary chat work in that chat's project; they are not isolated research tasks.