From c1f187a5ad4eb04059acf50d7ffffd722b26fb89 Mon Sep 17 00:00:00 2001 From: Michael Behr Date: Fri, 13 Feb 2026 17:00:29 -0500 Subject: [PATCH 1/2] Document agent participants. --- docs/_data/nav.yml | 2 + docs/researchers/add-agent-participant.md | 78 +++++++++++++++++++++++ 2 files changed, 80 insertions(+) create mode 100644 docs/researchers/add-agent-participant.md diff --git a/docs/_data/nav.yml b/docs/_data/nav.yml index 975898153..5e738b906 100644 --- a/docs/_data/nav.yml +++ b/docs/_data/nav.yml @@ -8,6 +8,8 @@ url: /researchers/run-experiment - title: Add LLM mediators to chat url: /researchers/add-mediator + - title: Add LLM agent participants to experiment + url: /researchers/add-agent-participant - title: Add Prolific integration url: /researchers/prolific - title: Developers diff --git a/docs/researchers/add-agent-participant.md b/docs/researchers/add-agent-participant.md new file mode 100644 index 000000000..c5830f16a --- /dev/null +++ b/docs/researchers/add-agent-participant.md @@ -0,0 +1,78 @@ +--- +title: Add LLM agent participants to experiment +layout: default +--- + +LLM agents can join experiments as participants. Each agent participant can have +its own individual prompt, but will otherwise run through an experiment in the +same way as a human would. + +## Preparing an Experiment for Agent Participants + +You can add agent participants to any experiment, as long as you have a Google +API key configured. No other experiment-level config is necessary. However, we +recommend setting up your experiments with an eye for how agent participants +will see each stage. + +**Experiment info**: Agent participants will see any text in the experiment info + stage, but they won't see the contents of a linked Youtube video. + +**Stage metadata**: This is where agent participants will see what each stage is + about, so consider how clear your stage names and instructions are. + +**Progress settings**: Agent participants may move through your experiment + faster than you expect, or get stuck on chat stages where you don't expect. + Consider checking "Wait for all active participants to reach this stage before + allowing progression" before discussion stages, to prevent agents from moving + on from a chat before humans arrive. Also consider setting a chat time limit, + or describing specific goals for each chat stage. + +**Profile settings**: If you select the option to assign random animal profiles + to participants, be aware that the chosen animal could influence the agent's + behavior! The agent will be reminded in its profile prompt that it is a human + and not actually the given animal, so most models shouldn't try to respond as + though they were the animal, but we can't rule out subtler effects. + +## Adding Agent Participants to a Cohort + +To add agent participants to a cohort: + +- From the experiment overview screen, hit the icon to add a participant, and + select "Add agent participant" from the menu. +- From the cohort management screen, click the icon at the top of the "Agent + participants" section, which should appear between human participants and + agent mediators. + +You'll see a window to configure the agent: you'll need to select a +model for the agent to use, and you can optionally add a prompt context to give +to the agent. The prompt context may be useful for e.g. giving different +personalities or instructions to different agents. These settings will apply for +that agent across all experiment stages. + +## Supported Stages + +- Terms of Service +- Info +- Set Profile: If allowed to set their profile, agents will usually choose based + on their prompt context. +- Survey +- Group Chat +- Private Chat +- Survey / Survey Per Participant +- Ranking + +Not currently supported: + +- Role assignment +- Comprehension check +- Payout +- Reveal + +## Debugging Agent Participants + +We recommend always doing test runs with agent participants before launching +your experiment. To see the details of an agent participant's response, click +the "LLM Logs" button on the left sidebar. Even if the agent is responding as +you expect, we recommend reviewing the prompts sent to the agent at each stage, +at least once. This will help you confirm that the agent is seeing exactly the +information it should be. From a44fca7a0dc68193fd0feff93fe04874763feac7 Mon Sep 17 00:00:00 2001 From: Michael Behr Date: Fri, 27 Feb 2026 15:56:32 -0500 Subject: [PATCH 2/2] Clarify agent participant model config and point to implementation doc --- docs/researchers/add-agent-participant.md | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/docs/researchers/add-agent-participant.md b/docs/researchers/add-agent-participant.md index c5830f16a..80f3f4a02 100644 --- a/docs/researchers/add-agent-participant.md +++ b/docs/researchers/add-agent-participant.md @@ -10,7 +10,8 @@ same way as a human would. ## Preparing an Experiment for Agent Participants You can add agent participants to any experiment, as long as you have a Google -API key configured. No other experiment-level config is necessary. However, we +API key configured. (Right now, only Gemini models can be selected for agent +participants.) No other experiment-level config is necessary. However, we recommend setting up your experiments with an eye for how agent participants will see each stage. @@ -60,6 +61,8 @@ that agent across all experiment stages. - Private Chat - Survey / Survey Per Participant - Ranking +- Asset Allocation +- Stock Info Not currently supported: @@ -68,6 +71,9 @@ Not currently supported: - Payout - Reveal +For details on how a stage implements agent participants, see +[Add stage](../developers/add-stage). + ## Debugging Agent Participants We recommend always doing test runs with agent participants before launching