Devin records an iPhone game test
Cognition reports Devin testing the iPhone game Otter Run in a simulator and returning a recording with checks marked passed or untested for engineers to inspect.
openai.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Test this authorized iPhone game in its simulator. Record the session, report each observed check and clearly mark untested behavior so an engineer can review the evidence.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Walden Yan via OpenAI customer account
- Original account
- Read the source ↗
- Published
- 11 Sept 2026
- Source link checked
- 15 Oct 2026
- Reported result
- Cognition reports Devin testing the iPhone game Otter Run in a simulator and returning a recording with checks marked passed or untested for engineers to inspect.
- Human involvement
- The recording and scoped report are available for engineers to inspect. The account does not describe a completed human review or exhaustive test coverage.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Cognition gives Devin the Otter Run iPhone game for testing with GPT-6 Astra.
- Devin runs the application in an iPhone simulator and records behavior.
- It returns the recording and a report distinguishing passed checks from untested areas.
Cognition's cofounder Walden Yan describes using Devin with GPT-6 Astra to test software and return evidence. The concrete example in the OpenAI customer account is Otter Run, an iPhone game. Devin runs the game in a simulator and returns a recording alongside a test report.
The recording lets engineers inspect actual application behavior. The report identifies checks that passed and areas left untested, preserving the difference between an observed success and coverage that was never attempted. The model is part of Devin's execution here; the agent should be attributed to Devin rather than to Codex or ChatGPT agent.
The same account describes customer bug screenshots being handed to Devin and repaired with a screenshot of the result. That is broader context for the evidence-returning workflow, not a second completed Otter Run task.
The source does not name particular game mechanics, report an exhaustive test suite or independently verify that the recording proves every expected behavior. Its supported outcome is a simulator recording and a scoped report that engineers can review. The prospect of less manual code review is a company expectation, not a measured reduction established by this example.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: openai.com · Walden Yan via OpenAI customer account
Frequently asked questions
Which agent performed the task?
Devin, using GPT-6 Astra as its model.
What evidence was returned?
A simulator recording and a report with passed checks and untested areas.
Did the source establish complete coverage?
No. The report explicitly includes areas left untested.
More details
Remaining steps
- Devin returns artifacts that engineers can inspect to understand observed behavior and remaining scope.
About this account
Devin operated the iPhone simulator, recorded its game interaction and returned a report distinguishing passed and untested items.
Firsthand account opened and read. Outcome is source-reported, not independently reproduced. Suggested prompt written by Agents In Use. Customer interview is primary quoted account; recovered by web after local403. No gameplay accuracy or coverage claim beyond published evidence.
Credit: Walden Yan via OpenAI customer account
Source published 11 Sept 2026
Source checked 15 Oct 2026
Tools and topics
Devin · iPhone simulator · GPT-6 Astra
Devin · iPhone simulator · GPT-6 Astra
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.