# Plan and execute a task

Turn the requested result into supported operations and adapt when the current route needs recovery.

A task starts with the result you want, not a list of tool calls. The agent should choose operations that can produce that result with the account, data, and constraints you supplied.

For example, “summarize the current project decisions” calls for evidence and synthesis. “Update the project document” also needs the exact document, the intended change, and a readback of what was saved.

## Make the result concrete

Before acting, establish:

- The output or changed state.
- The relevant project, account, resource, and time range.
- Constraints that affect the work, including exclusions.
- The observation that will establish completion.

Use facts already present in the conversation or authorized sources. Ask a focused question only when an unresolved detail changes the result or a consequential destination. Routine tasks need a short sequence; larger tasks benefit from a visible plan showing dependencies and deliverables.

## Choose an executable route

| Needed result | Starting route |
| --- | --- |
| Find information already retained in the Vault | [Search saved information](/docs/vision/tools/search). |
| Read current service data or change an app object | [Run app actions](/docs/vision/tools/actions). |
| Operate a webpage | [Use a browser](/docs/vision/browser-computer/browser-use). |
| Work in a desktop application | [Use your computer](/docs/vision/browser-computer/computer-use). |
| Read or deliver a local file | [Work with local files](/docs/vision/tools/local-files). |

Inspect the current source and method before building an invocation. For example, a connected Gmail source can be checked and read through the terminal:

```sh
indra --root "$indra_root" manage --json '{"what":"sources","about":"gmail"}'
indra --root "$indra_root" action --json '{"name":"connector.read","params":{"source":"gmail","request":{"limit":5}}}'
```

Here `$indra_root` is your existing Vault path. The read returns a bounded current result through the local Action path; it does not promise that the returned content is retained for later SQL search. The exact method for another provider may differ, so use [tool discovery](/docs/vision/tools/discovery).

## Follow the results in order

A practical execution loop is:

1. Inspect the current source, object, page, or file.
2. Select the next operation and its current arguments.
3. Execute it once.
4. Inspect the returned result and the changed state where relevant.
5. Continue from that result, or take the indicated recovery step.

Use returned object IDs, session IDs, and schemas. Do not invent a destination or reuse a stale browser control. Independent reads can proceed together; a write that depends on account selection, current content, or a completed preview must wait for that prerequisite.

When multiple agents help, give each a defined scope. Keep one owner for a shared edit or transaction so two callers do not apply the same change.

## Keep recovery inside the task

| What happened | Useful next step |
| --- | --- |
| A needed source is not connected | Continue its supported [connection recovery](/docs/vision/connectors/recovery), then return to the task. |
| A method or argument is rejected | Read the current installed schema and correct the invocation. |
| A job is pending | Follow the returned job identity; see [Long-running work](/docs/vision/autopilot/long-running-tasks). |
| A write lost its response | Read the destination before retrying. |
| A required credential or device is unavailable | Identify that exact missing step and preserve other completed work. |
| An available capability cannot meet the goal | Try a capable authorized alternative, or state the specific implementation gap. |

A recovery page is part of the existing connection transaction. Keep its owner alive and avoid starting a competing sign-in. Routine account selection and consent should remain agent work when available tools can complete them. Explicit opt-outs, denials, and cancellation remain binding.

## Finish at the requested result

After a connection or sync completes, continue the research, file, or app task that required it. Close the task when its result is verified, the user cancels, or a concrete blocker leaves no useful continuation.

Do not expand a handbook edit into product implementation or a package release automatically. The requested delivery stage determines what remains to be done. See [Review and deliver results](/docs/vision/autopilot/review-and-delivery).
