Replay part of one of your own Claude Code or Codex conversations with three other coding agents, and see how it would have gone.
You already sent sessions from this browser.Open your results page to continue where you left off.
This study needs a laptop or desktop computer with Chrome, Edge or Safari: the computer where your Claude Code or Codex sessions are recorded. It does not work on a phone or tablet.
What this study does
Think of one of your conversations with Claude Code or Codex. How would it have gone with a different coding agent: Codex instead of Claude Code, or the same agent running on a larger or cheaper AI model, such as Opus instead of Haiku?
In this study we find out with one of your own past conversations, which we call a session. Claude Code and Codex save each session as a file on your computer. From your session, we create a simulated you: a program that writes to a coding agent the way you do. Then we pick one of your messages in the middle of the session and replay what came after it with three different coding agents. The simulated you writes to each agent in your place, asking for what you asked for in the original.
You read the three replays and tell us which agent you would rather have had. You also tell us whether the simulated you really writes like you, since the replays are only as good as it is.
This study involves the following steps
Eligibility check in your browser (a few minutes, longer for a large folder). Not every session can be replayed: we need one where, partway through, you corrected or pushed back on the agent, because that is where different agents go different ways. Your own browser looks through your session files on your computer for such sessions, and nothing leaves it. Your browser will ask to "upload" the folder: that only lets the check run in your browser. If none can be used, you return to Prolific and are paid £0.50.
Consent, then choose. If some can be used, you read and agree to the consent. Then you see exactly what would be sent from each usable session and choose which, if any, to send.
Background questions, then a wait. Six short questions, then 20 to 40 minutes while we create the simulated you and run the three replays. You can do something else meanwhile.
Compare the three replays. First you predict which agent will do best. Then you read the replays side by side, mark each message from the simulated you as like you or not, and rank the agents.
At the end, we reveal which agent was which: the coding agent and AI model behind each, and what each one cost.
The whole study takes about 75 minutes, including the wait.
We pay you £11 through Prolific for finishing the study.
If none of your sessions can be used, or you decide not to send any, we pay you £0.50 for the check.
If you withdraw after sending your sessions, we pay you £3.
Research team
Led by researchers at the KAIST School of Computing (Korea Advanced Institute of Science and Technology). Approved by the KAIST Institutional Review Board (KAISTIRB-2026-189).
This link does not include a Prolific ID or an invite code, so we cannot match you to the study or pay you. If you came from Prolific, open the study again from its page on Prolific. If the researchers invited you, open the whole link they sent you, or write to taesoo.kim@kaist.ac.kr.
Start the eligibility check
First, your browser looks through your Claude Code or Codex sessions for ones that can be replayed: sessions where, partway through, you corrected or pushed back on the agent. If you came from Prolific, we record your Prolific ID and how the check went (counts and reasons, never what your sessions contain). This record is deleted 2 years after the study ends.
Taking part is voluntary. KAIST IRB approval KAISTIRB-2026-189.
Which tool's sessions should we check? If you use both, choose the one you've used more.
Follow these steps to start the check:
1. Copy the path of your sessions folder.
2. Press Choose your sessions folder below. In the folder window that opens:
Your browser then shows a button named Upload and asks whether to "upload" the files to this site. Press it: that's its word for letting this page read the folder.
Nothing is sent to our server during the check: your browser reads your session files on your computer, in this tab.
Tick the box at the top of this card first.
Starting…
We check every session in the folder. This can take a few minutes, so keep this tab open.
You chose not to send any sessions, so nothing was sent. You can close this tab.
Payment for the check
We compensate you £0.50 for the check, paid through Prolific. Press Return to Prolific to be paid. After that, you can close this tab.
None of the session files in the folder you chose can be used, so you can't take part with them.
A session can be used when the conversation has enough back-and-forth after one of your messages, and the files it worked on can be recreated.
You can close this tab.
If you chose the wrong folder, or your sessions are in the other tool, check again:
Payment for the check
We compensate you £0.50 for the check, paid through Prolific. Press Return to Prolific to go back and be paid.
What happens next
Plan for about 75 minutes in all: about 35 minutes of your own time, and a 20 to 40 minute wait while the agents run. During the wait you can do something else: leave the study tab open in the background and it notifies you when the conversations are ready, or close it and come back with your results link.
Informed consent agreement(about 2 minutes). Read what we collect, who reads it and how long we keep it, and agree.
Choose sessions(about 3 minutes). See exactly what would be sent from each usable session, tick the ones you're comfortable sending, and send them.
Background questions, then a wait(about 2 minutes, then 20 to 40 minutes). Meanwhile, we create the simulated you from your session and run the three replays.
Compare the three replays(about 20 minutes). You see where your session was and the message the replays start from. You're told which three agents took part, but the replays are shown without names, and you first predict which agent will handle your message best. Then you read the three replays side by side, round by round, and mark each message from the simulated you as like you or not. Finally you rank the agents and rate how much the simulated you writes and reacts like you, saying why in a sentence each.
At the end, we reveal which agent was which(about 3 minutes): the coding agent and AI model behind each, what each one cost, and your prediction beside your ranking. Then two short questions ask what you take away from it.
We pay you £11 through Prolific when you finish. The consent page lists every payment.
Your sessions are sent to our server only to create the simulated you and run the three conversations. Once the conversations are ready, we delete your sessions and the simulated you.
From the session we use, we keep your first message, a short summary of what happened before the message where the agents start, that message, a short summary of what you asked for next, and the three conversations (what the agents and the simulated you wrote, and what the agents did in the copy of your project).
Your prediction, your marks on the simulated you's replies, your ratings, your ranking and your written responses.
Your answers to the background questions.
Your Prolific ID.
The consent page says who reads this and how long we keep it.
Select each session you're comfortable sending. We replay one of them. Our server makes one more check that only it can make, so sending more than one helps if the first can't be used. Sessions we don't use are deleted once the replays are ready.
What would be sent
Only part of each session: everything from its start up to a few exchanges after the message where the replay would start, one where you corrected or pushed back on the agent. An exchange is one of your messages and the agent's reply. In a short session, that can be all of it.
Before anything is sent, our website deletes sensitive details from your sessions, including keys, passwords, your user name and email addresses, and puts a placeholder in their place. Two things are not deleted: your project folder's name, and people's names written in messages or code.
Each session below shows its first message and the sensitive details that will be deleted from it. Press Read what would be sent to see the full text of the conversation before you select a session.
About this study
Development of an Agentic Coding Session Simulator and Its Validation Through User Studies, run by Juho Kim (principal investigator) and Tae Soo Kim at the KAIST School of Computing, and approved by the KAIST Institutional Review Board (KAISTIRB-2026-189). This study replays part of one of your sessions with three different coding agents, using a simulated version of you that writes to each agent in your place. It asks which agent you would rather have had, and how closely the simulated you resembles you. The parts below say what you would send, what happens to it, and your rights.
1. What is sent, and why
From each session you choose to send: everything from its start up to five exchanges after the last message of yours the agents could start from. In the session view, these messages are marked "The agents may start from this message". That means your messages, the agent's messages, every command it ran and what it printed, and the contents of every file it read or wrote in that part, with the sensitive details in part 2 deleted. It is sent to our server only to create the simulated you and run the three conversations.
If you came from Prolific, we also keep your Prolific ID with what you send, so that we can pay you. If you withdraw, we delete your sessions, the conversations and your answers, but keep your Prolific ID and which completion code you were given, so that you can still be paid and can't take part twice. We delete your Prolific ID 2 years after the study ends.
2. Sensitive details deleted before sending
Before anything is sent, our website deletes these sensitive details and puts a placeholder in their place:
text shaped like a key, token or password, with a note saying what it was, such as "[AWS access key removed by the study]"
your user name with "participant"
email addresses with person1@example.com, person2@example.com and so on
computer names with host1, host2 and so on, and network addresses with 192.0.2.1, 192.0.2.2 and so on
the folders between your home folder and your project folder with folder1, folder2 and so on
account names in GitHub and GitLab addresses with account1, account2 and so on
names in git author lines with Person 1, Person 2 and so on
Each detail gets the same placeholder everywhere, so the agents' copy of your files still matches the session. Our server deletes them again when the sessions arrive.
Not deleted: your project folder's name, because the code and its imports use it. People's names written in messages or code are not deleted either, because our website can't recognise names in text.
3. What the check can miss
The deletion only finds text that matches fixed patterns. A session where nothing was deleted can still contain a password, a name, or code you're not allowed to share. You decide what to send: on the next page, press Read what would be sent to read all of a session before you select it.
4. Who reads it
We send the sessions to OpenAI and to Amazon Web Services, which runs Anthropic's Claude models. Both process them in the United States.
An OpenAI model reads whether you pushed back on or corrected the agent at the marked message. We use one session where you did, and the models create the simulated you from it.
Three coding agents then continue that session from that message. Each works on a copy of the files the session showed, in its own isolated container that can reach only its model provider.
OpenAI doesn't use this data to train its models, and keeps it for up to 30 days to detect abuse. Amazon Web Services doesn't store it and doesn't share it with Anthropic.
The researchers running the study read what we keep (part 5) and your answers.
5. How long we keep it
Once the three conversations are ready, we delete everything you sent and everything made from it, including the simulated you. We keep, for 2 years after the study ends, your first message, a short summary of what happened before the message where the agents start, that message, a short summary of what you asked for next, and the three conversations (what the agents and the simulated you wrote, and what the agents did in the copy of your project), and your answers; then we delete them. If our server stops with an error, we keep what you sent until we have looked at the error, then delete it. We use them only for this research. Apart from the model providers in part 4, we don't share them with anyone outside the research team, and we never publish or release them as a dataset. We publish only the study's results: figures, and short excerpts with nothing that identifies you. Before publishing an excerpt, we remove names, user names, email addresses and file paths from it. We delete them sooner whenever you ask.
6. Your rights
Taking part is voluntary. You can stop at any time without giving a reason and without any disadvantage. Until you finish the study, you can withdraw with the Withdraw and delete my data button at the top right of every page after you send your sessions. After you finish, contact the researchers (taesoo.kim@kaist.ac.kr) to have your data deleted.
If you withdraw before finishing, we pay you £3 as a bonus through Prolific. If you ask us to delete your data after finishing, you keep your £11.
Withdrawing deletes from our server everything you sent, the three conversations and your answers. You can also contact the researchers to have them deleted.
Copies the model providers already received are kept as part 4 says.
If you withdraw while the agents are running (up to about 25 minutes of the wait), your withdrawal counts at once and we delete your data as soon as they end.
If you came from Prolific and stop before finishing without withdrawing, we don't pay you, with two exceptions. If the eligibility check finds no usable session, we pay you £0.50 for the check. If our server finds after you send your sessions that none can be used, we pay you £3 as a bonus.
For questions about your rights as a participant, contact the KAIST IRB office (+82-42-350-2189). The KAIST Institutional Review Board approved this study (KAISTIRB-2026-189).
7. Risks and benefits
The main risk is that something private in your session is sent to our server and the model providers while the simulations run. Three things reduce it: sensitive details are deleted before sending, you choose which sessions to send after reading them, and we delete your sessions once the conversations are ready. What we keep (part 5) can still include text from your session, such as your first message and what the agents read in the copy of your files. You receive the payment above, and you see how three different coding agents handle a moment from your own work.