---
title: "Your agent says done: here is how to check"
description: "An agent saying done is a claim. How to check it in Lusor: the video recording, the step-by-step replay with a screenshot after each action, the Replay page's exports, the recording and replay API endpoints, and the limits of a recording."
canonical: "https://lusor.ai/blog/your-agent-says-done-how-to-check"
updated: "2026-11-15"
---

[Blog](https://lusor.ai/blog) / [Insights](https://lusor.ai/blog/insights)

# Your agent says done: here is how to check

When an agent replies that the task is done, that reply is a claim. Lusor keeps the evidence beside it: a video of the phone, every step with the screenshot taken after it, and an API that returns both. Here is how to use them, and what they cannot tell you.

Recordings · November 15, 2026 · 9 min read

## TL;DR

- An agent's reply that a task is done is a claim. Lusor keeps a record of every task so the claim can be checked against what happened on the phone.
- Turns sent to an agent are always recorded on video. Sessions started without an agent are recorded only when the request sends recording: true.
- The step-by-step replay shows each action the agent took, with the screenshot taken after it. The Replay page adds a waterfall timeline, a playhead you can drag and the step JSON.
- The API returns the same evidence: /v1/sessions/{id}/recording for the video's MP4 segments, /replay for the actions and their screenshots, and /replay/frame for one screenshot as a PNG.
- A recording shows the screen, not whether the result is right, and it pauses while a secret is typed. On Free, history and video are kept for 7 days.

An agent finishes a task and replies that it is done. The message was sent, the booking was made, the forty listings were read. That reply is a claim, and the cases where it is wrong are the ones that cost something: the message went to the wrong thread, the booking is for the wrong day, the list stopped at thirty.

Lusor keeps a record of every task so you can check what the agent did. What follows is what that record holds, how to read it in the app and through the API, and what it cannot tell you.

## Why is “done” not enough?

Because the reply is the agent's own account of its work, and the agent can be wrong. A summary written at the end of a run leaves out the same things the run got wrong: an agent that misread a screen will describe the task as it understood it.

Checking the phone afterwards helps, but only with the end state. It shows that a message exists, not which draft was discarded on the way, and it shows nothing at all when the task was to read something. What settles the question is a record made while the work happened, by something other than the agent's summary.

## What does Lusor record while an agent works?

Lusor records the messages, the list of steps the agent took, a screenshot of the phone for every step, and a video of the phone screen while the agent worked. The [recordings and history guide](https://docs.lusor.ai/help/recordings-and-history) lists each part.

- **Messages:** your messages and the agent's replies.
- **Steps:** the list of steps the agent took, shown in the chat.
- **Screenshots:** a screenshot of the phone for every step.
- **Video:** a recording of the phone screen while the agent worked. This works on your own iPhone or Android phone and on Lusor Phones.
- **Detailed step breakdown:** the full technical record of every step, shown in the replay. This is included on paid plans.

Whether there is a video depends on how the session started. Turns sent to an agent are always recorded. [Sessions without an agent](https://docs.lusor.ai/agents/send-a-turn#sessions-without-an-agent), which are started on a phone directly through the API, are recorded only when you ask: send `"recording": true` when you start the session. It is off by default.

## How do you watch a run back in the app?

Hover over a finished reply from the agent and choose Replay. The replay plays the phone recording next to the agent's steps, and you can move through it with the scrubber.

The [Replay page](https://docs.lusor.ai/changelog) is built for looking closely. It has a waterfall timeline, a playhead you can drag, and the step JSON beside it. Two more things on it matter when the check has to leave the app.

- **Export the results.** A run's results export as CSV, TSV, JSON or Markdown, so rows an agent collected can be compared with the screens they came from. That is the check to run on a table before it is [saved to a Sheet](https://lusor.ai/blog/stop-copying-app-data-by-hand) that other work depends on.
- **Save a frame.** A single frame saves as an image, which is enough to show someone else the screen a step ended on.

## What does the step-by-step replay show?

The [step-by-step replay](https://docs.lusor.ai/agents/recordings-and-replay#step-by-step-replay) shows each action the agent took, with the screenshot taken after it. The video answers what the screen looked like over time. The steps answer what the agent did and what each action led to.

That pairing is what makes a check quick. A reply that says a form was submitted can be compared with the screenshot after the last tap: a confirmation screen, or the same form with an error under one field. A reply that says a setting was changed can be compared with the screenshot after the toggle. There is no need to watch the whole video to find the moment, because every action already has its own picture.

A check that takes a minute or two goes in this order.

- Read the reply and pick out what it claims: what was done, to what, and with which values.
- Open the replay and go to the last steps. Compare the final screenshots with the claim.
- For anything that was typed or chosen, find the step where it happened and read the screenshot after it.
- For a list, count the rows and compare a sample with the screenshots they came from, and all of them when a decision rests on the numbers.

## How do you check a run through the API?

With three endpoints under the session: `/v1/sessions/{id}/recording` for the video, `/v1/sessions/{id}/replay` for the actions and their screenshots, and `/v1/sessions/{id}/replay/frame` for one screenshot. The [recordings and replay reference](https://docs.lusor.ai/agents/recordings-and-replay) has examples in three languages. This is the request for a recording:

`curl https://api.lusor.ai/v1/sessions/<session id>/recording -H "X-Api-Key: $LUSOR_API_KEY"`

The response always has `recording_status`, which takes one of four values.

|  | Meaning |
| --- | --- |
| absent | The session has no recording. |
| pending | The recording is still being finished. Ask again in a few seconds. |
| complete | The recording is finished. |
| failed | The recording could not be finished. recording_status_reason says why. |

`data` is `null` until there is something to play. Then `data.segments` lists the recording's MP4 segments in order. Each segment has a `url`, its `start_offset_ms` and `duration_ms` within the recording, and `expires_at`, the time in Unix seconds when the `url` stops working. Request the endpoint again for fresh links.

`/replay` lists each completed action in time order. Each entry has the action `event`, its `timestamp`, and a `frame_url` for the PNG screenshot taken after it. The link expires about five minutes later, and screenshots that are still uploading appear on a later request. To download one screenshot through the API instead of a link, request `GET /v1/sessions/{id}/replay/frame?index=<n>`, where `index` counts from 0 in the list `/replay` returns. The response is the PNG image.

For code that checks its own runs, the [session artifacts](https://docs.lusor.ai/agents/follow-a-session#artifacts) endpoint returns the session's output with a list of its screenshots, each with its step and time. A script can store the output next to the last frame, so a person reviewing it later sees the claim and the screen together.

## Where can you see recordings of real runs?

On the [AndroidWorld benchmark page](https://lusor.ai/benchmarks/androidworld), which publishes video and steps for all 116 tasks of Lusor's run. Each task opens its recording and steps.

That is the same idea applied to a benchmark score. A pass rate is a claim too, and a published recording of every task lets a reader open the ones they doubt, including the tasks where the run was slower or a grader disagreed.

## What are the limits?

A recording shows what was on the phone's screen, and nothing else. That is a lot, but it is not everything, and the record has gaps and an end date.

- **It shows the screen, not the outcome elsewhere.** A confirmation screen shows that the app said the order went through. Whether the order reached the shop is outside the phone, and no recording of the phone shows it.
- **It shows what was done, not whether it was the right thing.** A recording of an agent messaging the wrong contact is a complete and accurate recording. The record makes the mistake findable. It does not flag it.
- **It pauses for secrets and questions.** The recording pauses while the agent waits for you to answer a question, and while a secret, such as a password, is being typed on the phone. The replay marks those gaps as Parked, waiting for human and Secret window. On iPhone, a typed Vault secret is masked in the run's screenshots and events. What else an agent can see on a phone, and who can open the record, is covered under [phone agent safety](https://lusor.ai/blog/is-it-safe-to-let-an-agent-use-your-phone).
- **Sessions without an agent are not recorded by default.** A session started on a phone directly through the API has a video only when the request sent `"recording": true`. Without it, `recording_status` is `absent`.
- **History has a window.** Recordings and replay follow the plan's history window. Messages, steps and screenshots are kept for 7 days on Free and without a limit on paid plans. Video replays are kept for 7 days on Free and 30 days on paid plans. For a session older than the window, both endpoints return `402` with `error: "upgrade_required"` and `reason: "history_window"`. A recording stays available past the window when the plan includes video replay, or when it was made on a Lusor Phone with a paid plan.
- **Some of the record depends on the plan.** The detailed step breakdown, the full technical record of every agent step as JSON, is included on Hobby and above. Every plan sees the step list in the chat. Paid plans include minutes of screen recording each month, starting at 2,000 on Hobby. Free plan sessions are still recorded and replayable for 7 days without a minutes pool. [The pricing page](https://lusor.ai/pricing) has the full table.
- **Links expire.** A segment's link and a frame's link both stop working at their `expires_at`. Save the file, not the link, when a recording has to outlast the session's window.

None of this makes the record less worth having. It sets what to expect from it: proof of what the agent did on the phone, in order, with the screen after each action. The next time an agent says done, open the replay and look at the last three steps before taking its word.

## Frequently asked questions

### Is every Lusor session recorded on video?

Turns sent to an agent are always recorded. Sessions started without an agent are recorded only when the request that starts the session sends recording: true, which is off by default.

### How do you watch a run back?

Hover over a finished reply from the agent and choose Replay. The replay plays the phone recording next to the agent's steps. The Replay page has a waterfall timeline, a playhead you can drag, and the step JSON beside it.

### What does the step-by-step replay show?

Each action the agent took, with the screenshot taken after it. Through the API, the replay endpoint lists each completed action in time order with the action's event, its timestamp and a link to the PNG screenshot taken after it.

### Can a recording be fetched through the API?

Yes. A GET request to /v1/sessions/{id}/recording returns recording\_status and, when there is something to play, the recording's MP4 segments in order. Each segment's link expires, and requesting the endpoint again returns fresh links.

### How long are recordings kept?

Recordings and replay follow the plan's history window. Messages, steps and screenshots are kept for 7 days on Free and without a limit on paid plans. Video replays are kept for 7 days on Free and 30 days on paid plans.

### What does a recording not prove?

It shows what was on the phone's screen, not what happened anywhere else, and it does not say whether the agent picked the right thing to do. It also has gaps: the recording pauses while the agent waits for an answer and while a secret is being typed on the phone.

## Keep reading

- [Stop copying app data by hand: have an agent build the sheet](https://lusor.ai/blog/stop-copying-app-data-by-hand): Useful data often lives only inside phone apps, with no export and no API. How a Lusor agent reads it on a phone and saves it as rows in a Sheet, how the rows get out as CSV, TSV, JSON or Markdown, and what to check before relying on them.
- [What makes a phone agent fast](https://lusor.ai/blog/what-makes-a-phone-agent-fast): Lusor's median is 48.2 s per AndroidWorld task, against 216.9 s for Mobilerun's published times (4.5×). Where a phone agent's time goes, what makes Lusor's runs fast, the 5 tasks where it was slower, and the trade-offs with accuracy and cost.
- [Is it safe to let an agent use your phone?](https://lusor.ai/blog/is-it-safe-to-let-an-agent-use-your-phone): What a Lusor agent can see on your phone, where saved logins and secrets live, how long screenshots and recordings are kept, who in a workspace can view a session, and the limits to accept before connecting a phone.
