Using EDSL for AI polling

The sections below include:

Creating an AI agent: Basic steps to construct an AI agent.
Administering questions: How to create questions and prompt agents to answer them.
Selecting language models: Specify language models that you want to use to generate responses.
Analyzing results: Examples of built-in methods for analyzing responses as datasets.
Designing agent traits: How to construct agents with complex personas and traits.
Converting surveys into EDSL: Import other surveys into EDSL to analyze and extend them with agents.
Constructing agents from survey data: Use survey responses to construct agents representing respondents.

Sample data: Cooperative Election Study Common Content, 2022

For purposes of demonstration, we use data from the Cooperative Election Study Common Content, 2022 in several ways:

In this notebook we use lists of respondent attributes from the Breakdown of National Vote for U.S. House (CES validated voters) (CES Guide 2022 pp.24-25) to design agents with combinations of the attributes, and then administer questions to them.

Companion notebooks:

In Working with CES data in EDSL we reconstruct the 2022 CES Pre-Election Questionnaire in EDSL. The formatted questions can be modified and administered to agents in new surveys as desired. The notebook also contains dictionaries of the CES questions and answer codebook for convenience in working with the data.
In Creating agents for survey respondents we use the dataset of responses to the 2022 CES Pre-Election Questionnaire to create agents representing the respondents, and administer follow-on questions with them.

Reference & contact

Documentation for the EDSL package is available at https://docs/expectedparrot.com. You can also find example code, tutorials and notebooks for a variety of use cases. Please let us know if you have any questions or encounter issues working with this data:

Discord: https://discord.com/invite/mxAYkjfy9m
Email: info@expectedparrot.com

Technical setup

EDSL is compatible with Python 3.9-3.12. See instructions on installing the EDSL library and storing API keys for the language models that you want to use. In examples below where no model is specified, EDSL will use GPT 4 by default (an API key for OpenAI is required). We also show how to use different models.

Creating an AI agent

In this section we show how to create an AI agent and give it desired attributes. For more details on constructing and using AI agents please see our documentation page on agents. We start by importing the tools for creating agents:

from edsl import Agent

Here we create a simple Agent and pass it a dictionary of traits. We optionally include a narrative persona and also specify traits individually for use in segmenting and analyzing survey responses:

agent = Agent(
    traits={
        "persona": "You are 55-year-old research scientist living in Cambridge, Massachusetts.",
        "occupation": "Research scientist",
        "location": "Cambridge, Massachusetts",
        "age": 55,
    }
)

We can access the traits directly:

agent.location

'Cambridge, Massachusetts'

Designing agent panels

We can also create panels of agents in an AgentList and administer surveys to all of the agents at once. Here we construct combinations of traits from lists of respondent attributes in the CES Guide (see source details above). (Information can be imported from a variety of data source types; see documentation for details.)

sex = ["Male", "Female"]
race = ["White", "Black", "Hispanic", "Asian", "Other"]
age = ["18-29", "30-44", "45-64", "65 and over"]
education = [
    "High school or less",
    "Some college/assoc. degree",
    "College/graduate",
    "Postgraduate study",
]
income = [
    "Under $30,000",
    "$30,000 to $49,999",
    "$50,000 to $99,999",
    "$100,000 to $199,999",
    "$200,000 or more",
]
party_affiliation = ["Democrat", "Republican", "Independent/Other"]
political_ideology = ["Liberal", "Moderate", "Conservative", "Unsure"]
religion = [
    "Protestant/other Christian",
    "Catholic",
    "Jewish",
    "Something else",
    "None",
]
evangelical = ["Yes", "No"]
married = ["Yes", "No"]
lgbt = ["Yes", "No"]

Here we create a method to generate a list of agents with randomly selected combinations of traits:

from edsl import AgentList
import random

def generate_random_agents(num_agents):
    agents = []
    for _ in range(num_agents):
        agent_traits = {
            "sex": random.choice(sex),
            "race": random.choice(race),
            "age": random.choice(age),
            "education": random.choice(education),
            "income": random.choice(income),
            "party_affiliation": random.choice(party_affiliation),
            "political_ideology": random.choice(political_ideology),
            "religion": random.choice(religion),
            "evangelical": random.choice(evangelical),
            "married": random.choice(married),
            "lgbt": random.choice(lgbt),
        }
        agents.append(Agent(traits=agent_traits))

    return AgentList(agents)

Example usage:

num_agents = 3
agents = generate_random_agents(num_agents)

Agent instructions

If we want to give all the agents a special instruction, we can optionally pass an instruction to the agents (this can also be done when the agents are created):

for agent in agents:
    agent.instruction = "Today is July 1, 2022."

We can inspect the agents that have been created:

agents

AgentList agents: 3;

	sex	race	age	education	income	party_affiliation	political_ideology	religion	evangelical	married	lgbt
0	Female	Asian	30-44	College/graduate	$30 , 000 t o$ 49,999	Republican	Conservative	Jewish	Yes	No	Yes
1	Male	Hispanic	18-29	High school or less	$100 , 000 t o$ 199,999	Democrat	Unsure	Protestant/other Christian	Yes	No	Yes
2	Female	White	30-44	College/graduate	$100 , 000 t o$ 199,999	Republican	Liberal	Something else	Yes	Yes	No

Creating questions

An Agent is designed to be assigned questions to answer. In this section we construct questions in the form of Question objects, combine them into a Survey, administer it to some sample agents (from above), and inspect the responses in the dataset of Results that is generated. EDSL comes with many question types that we can select from based on the form of the response that we want to get back from the language model (free text, linear scale, checkbox, etc.). See examples of all question types. Here we create a multiple choice question from the CES Pre-Election Questionnaire (the response will be a selection from the list of options that we include) and compose a follow-up free text question (the response will be unstructured text):

from edsl import QuestionMultipleChoice, QuestionFreeText

# From the CES pre-election questionnaire
q_pid3 = QuestionMultipleChoice(
    question_name="pid3",
    question_text="Generally speaking, do you think of yourself as a ...?",
    question_options=["Democrat", "Republican", "Independent", "Other", "Not sure"],
)

# Potential follow-up question
q_views = QuestionFreeText(
    question_name="views", question_text="Describe your political views."
)

We combine the questions into a Survey to administer them together:

from edsl import Survey

survey = Survey([q_pid3, q_views])

Administering a survey

We administer a survey by calling the run method, after (optionally) adding agents with the by method:

results = survey.by(agents).run()

We can show a list of all the components of the Results that have been generated, and see that the results include information about the agents, questions, models, prompts and responses:

results.columns

	0
0	agent.age
1	agent.agent_index
2	agent.agent_instruction
3	agent.agent_name
4	agent.education
5	agent.evangelical
6	agent.income
7	agent.lgbt
8	agent.married
9	agent.party_affiliation
10	agent.political_ideology
11	agent.race
12	agent.religion
13	agent.sex
14	answer.pid3
15	answer.views
16	cache_keys.pid3_cache_key
17	cache_keys.views_cache_key
18	cache_used.pid3_cache_used
19	cache_used.views_cache_used
20	comment.pid3_comment
21	comment.views_comment
22	generated_tokens.pid3_generated_tokens
23	generated_tokens.views_generated_tokens
24	iteration.iteration
25	model.frequency_penalty
26	model.inference_service
27	model.logprobs
28	model.max_tokens
29	model.model
30	model.model_index
31	model.presence_penalty
32	model.temperature
33	model.top_logprobs
34	model.top_p
35	prompt.pid3_system_prompt
36	prompt.pid3_user_prompt
37	prompt.views_system_prompt
38	prompt.views_user_prompt
39	question_options.pid3_question_options
40	question_options.views_question_options
41	question_text.pid3_question_text
42	question_text.views_question_text
43	question_type.pid3_question_type
44	question_type.views_question_type
45	raw_model_response.pid3_cost
46	raw_model_response.pid3_one_usd_buys
47	raw_model_response.pid3_raw_model_response
48	raw_model_response.views_cost
49	raw_model_response.views_one_usd_buys
50	raw_model_response.views_raw_model_response
51	scenario.scenario_index

We can select and print components of the Results in a table (see examples of all methods for analyzing results):

results.select("age", "education", "pid3", "views")

	agent.age	agent.education	answer.pid3	answer.views
0	30-44	College/graduate	Republican	As an AI, I don’t have personal beliefs or political views. However, I can provide information on a wide range of political topics and help explain different perspectives. If you have any specific questions or need information on a particular political issue, feel free to ask!
1	18-29	High school or less	Democrat	As an AI, I don’t have personal beliefs or political views. However, I can provide information on various political ideologies, explain party platforms, or discuss current political events if you’d like. Let me know how I can assist you!
2	30-44	College/graduate	Republican	I don’t have personal political views or opinions. However, I can provide information on a wide range of political topics, discuss different political ideologies, and summarize various political perspectives based on data and research. If you have specific questions or topics in mind, feel free to ask!

Answer commentary

Question types other than free text automatically include a comment field for the agent to provide any unstructured commentary on its response to a question. This is useful in ensuring that responses are formatted as specified, providing an outlet for model verbosity. For example, in results.columns we can see that there is a field comment.pid3_comment. We can inspect this field as we do any other component of results. Here we also apply some pretty_labels to our table for readability:

(
    results
    .select("pid3", "pid3_comment")
    .print(
        pretty_labels={"answer.pid3": "Party", "comment.pid3_comment": "Comment"},
        format="rich",
    )
)

	Party	Comment
0	Republican	Given your traits, you have identified your party affiliation as Republican, which aligns with your political ideology.
1	Democrat	Based on your traits, you are affiliated with the Democratic Party, which aligns with the option “Democrat.”
2	Republican	Based on the provided traits, your party affiliation is listed as Republican.

Selecting language models

As mentioned above, if we do not specify a language model GPT 4 is used by default. We can also specify other language models to use in generating results, and compare responses for them. To see a list of all available models (uncomment the code):

from edsl import Model, ModelList

# Model.available()

To select models for a survey, pass the model names to Model objects:

models = ModelList(Model(m) for m in ["gemini-pro", "gpt-4o"])

We add a Model or list of models to a survey with the by method, the same as we do agents:

results = survey.by(agents).by(models).run()

results.select("model", "pid3", "pid3_comment")

	model.model	answer.pid3	comment.pid3_comment
0	gemini-pro	nan	Task failed with exception: Language model did not return a response for question ‘pid3.’.
1	gpt-4o	Republican	Given your traits, you have identified your party affiliation as Republican, which aligns with your political ideology.
2	gemini-pro	nan	Task failed with exception: Language model did not return a response for question ‘pid3.’.
3	gpt-4o	Democrat	Based on your traits, you are affiliated with the Democratic Party, which aligns with the option “Democrat.”
4	gemini-pro	nan	Task failed with exception: Language model did not return a response for question ‘pid3.’.
5	gpt-4o	Republican	Based on the provided traits, your party affiliation is listed as Republican.

Learn more about specifying language models.

Question context & memory

Survey questions are administered asynchronously by default, for efficiency. If we want an agent to have the context of one or more prior questions when presented a new question we can apply a rule specifying the questions and answers to add to the new question prompt:

from edsl import QuestionMultipleChoice, QuestionFreeText, Survey

# From the CES pre-election questionnaire
q_CC22_309e = QuestionMultipleChoice(
    question_name="CC22_309e",
    question_text="Would you say that in general your health is...",
    question_options=["Excellent", "Very good", "Good", "Fair", "Poor"],
)

q_CC22_309f = QuestionMultipleChoice(
    question_name="CC22_309f",
    question_text="Would you say that in general your mental health is...",
    question_options=["Excellent", "Very good", "Good", "Fair", "Poor"],
)

survey = Survey([q_CC22_309e, q_CC22_309f])

Here we add a memory of q_CC22_309e when administering q_CC22_309f and show the prompts that were administered:

survey = survey.add_targeted_memory(q_CC22_309f, q_CC22_309e)

results = survey.run()

results.select("CC22_309e_user_prompt", "CC22_309e", "CC22_309f_user_prompt", "CC22_309f")

	prompt.CC22_309e_user_prompt	answer.CC22_309e	prompt.CC22_309f_user_prompt	answer.CC22_309f
0	Would you say that in general your health is… Excellent Very good Good Fair Poor Only 1 option may be selected. Respond only with a string corresponding to one of the options. After the answer, you can put a comment explaining why you chose that option on the next line.	Good	Would you say that in general your mental health is… Excellent Very good Good Fair Poor Only 1 option may be selected. Respond only with a string corresponding to one of the options. After the answer, you can put a comment explaining why you chose that option on the next line. Before the question you are now answering, you already answered the following question(s): Question: Would you say that in general your health is… Answer: Good	Good

See examples of all methods for applying question context and memories (e.g., full memory of all prior questions, or a subset of questions).

Piping questions

We can also pipe individual components of questions into other questions. Here we use the answer to inputstate in the question text for CC22_320d:

from edsl import QuestionMultipleChoice, Survey

q_inputstate = QuestionMultipleChoice(
    question_name="inputstate",
    question_text="What is your State of Residence?",
    question_options=[
        "Alabama",
        "Alaska",
        "Arizona",
        "Arkansas",
        "California",
        "Colorado",
        "Connecticut",
        "Delaware",
        "District of Columbia",
        "Florida",
        "Georgia",
        "Hawaii",
        "Idaho",
        "Illinois",
        "Indiana",
        "Iowa",
        "Kansas",
        "Kentucky",
        "Louisiana",
        "Maine",
        "Maryland",
        "Massachusetts",
        "Michigan",
        "Minnesota",
        "Mississippi",
        "Missouri",
        "Montana",
        "Nebraska",
        "Nevada",
        "New Hampshire",
        "New Jersey",
        "New Mexico",
        "New York",
        "North Carolina",
        "North Dakota",
        "Ohio",
        "Oklahoma",
        "Oregon",
        "Pennsylvania",
        "Rhode Island",
        "South Carolina",
        "South Dakota",
        "Tennessee",
        "Texas",
        "Utah",
        "Vermont",
        "Virginia",
        "Washington",
        "West Virginia",
        "Wisconsin",
        "Wyoming",
    ],
)

q_CC22_320d = QuestionMultipleChoice(
    question_name="CC22_320d",
    question_text="Do you approve of the way the Governor of {{ inputstate.answer }} is doing their job?",
    question_options=[
        "Strongly approve",
        "Somewhat approve",
        "Somewhat disapprove",
        "Strongly disapprove",
        "Not sure",
    ],
)

survey = Survey([q_inputstate, q_CC22_320d])

results = survey.by(agents).run()

results.select("inputstate", "CC22_320d_user_prompt", "CC22_320d")

	answer.inputstate	prompt.CC22_320d_user_prompt	answer.CC22_320d
0	California	Do you approve of the way the Governor of California is doing their job? Strongly approve Somewhat approve Somewhat disapprove Strongly disapprove Not sure Only 1 option may be selected. Respond only with a string corresponding to one of the options. After the answer, you can put a comment explaining why you chose that option on the next line.	Somewhat disapprove
1	California	Do you approve of the way the Governor of California is doing their job? Strongly approve Somewhat approve Somewhat disapprove Strongly disapprove Not sure Only 1 option may be selected. Respond only with a string corresponding to one of the options. After the answer, you can put a comment explaining why you chose that option on the next line.	Somewhat approve
2	Florida	Do you approve of the way the Governor of Florida is doing their job? Strongly approve Somewhat approve Somewhat disapprove Strongly disapprove Not sure Only 1 option may be selected. Respond only with a string corresponding to one of the options. After the answer, you can put a comment explaining why you chose that option on the next line.	Somewhat disapprove

Survey rules & conditions

We can apply survey rules and conditions to administer relevant follow-up questions based on responses to questions. For example, here we add skip rules to a set of questions by calling the method add_skip_rule() and passing the target question and the condition to evaluate (questions not administered will show a None response):

from edsl import QuestionCheckBox, QuestionMultipleChoice, Survey

# From the CES pre-election questionnaire
q_CC22_300 = (
    QuestionCheckBox(  # Use checkbox to allow the agent to select multiple options
        question_name="CC22_300",
        question_text="In the past 24 hours have you... (check all that apply)",
        question_options=[
            "Used social media (such as Facebook or Youtube)",
            "Watched TV news",
            "Read a newspaper in print or online",
            "Listened to a radio news program or talk radio",
            "None of these",
        ],
    )
)

# Skip this question if the agent does not select "Watched TV news"
q_CC22_300a = QuestionMultipleChoice(
    question_name="CC22_300a",
    question_text="Did you watch local news, national news, or both?",
    question_options=["Local Newscast", "National Newscast", "Both"],
)

# Skip this question if the agent does not select "Watched TV news"
q_CC22_300b = QuestionMultipleChoice(
    question_name="CC22_300b",
    question_text="Which of these networks did you watch?",
    question_options=["ABC", "CBS", "NBC", "CNN", "Fox News", "MSNBC", "PBS", "Other"],
)

# Skip this question if the agent does not select "Read a newspaper..."
q_CC22_300c = QuestionMultipleChoice(
    question_name="CC22_300c",
    question_text="Did you read a print newspaper, an online newspaper, or both?",
    question_options=["Print", "Online", "Both"],
)

# Skip this question if the agent does not select "Used social media..."
q_CC22_300d = QuestionMultipleChoice(
    question_name="CC22_300d",
    question_text="In the past 24 hours, did you do any of the following on social media (such as Facebook, Youtube or Twitter)?",
    question_options=[
        "Posted a story, photo, video or link about politics",
        "Posted a comment about politics",
        "Read a story or watched a video about politics",
        "Followed a political event",
        "Forwarded a story, photo, video or link about politics to friends",
        "None of the above",
    ],
)

survey_CC22_300 = (
    Survey([q_CC22_300, q_CC22_300a, q_CC22_300b, q_CC22_300c, q_CC22_300d])
    .add_skip_rule(q_CC22_300a, "'Watched TV news' not in CC22_300")
    .add_skip_rule(q_CC22_300b, "'Watched TV news' not in CC22_300")
    .add_skip_rule(q_CC22_300c, "'Read a newspaper in print or online' not in CC22_300")
    .add_skip_rule(
        q_CC22_300d, "'Used social media (such as Facebook or Youtube)' not in CC22_300"
    )
)

results_CC22_300 = survey_CC22_300.by(agents).run()

results_CC22_300.select("CC22_300", "CC22_300a", "CC22_300b", "CC22_300c", "CC22_300d")

	answer.CC22_300	answer.CC22_300a	answer.CC22_300b	answer.CC22_300c	answer.CC22_300d
0	[‘Used social media (such as Facebook or Youtube)’, ‘Watched TV news’, ‘Read a newspaper in print or online’, ‘Listened to a radio news program or talk radio’]	National Newscast	Fox News	Online	Read a story or watched a video about politics
1	[‘Used social media (such as Facebook or Youtube)’, ‘Watched TV news’]	National Newscast	CNN	nan	Read a story or watched a video about politics
2	[‘Used social media (such as Facebook or Youtube)’, ‘Watched TV news’, ‘Read a newspaper in print or online’]	National Newscast	Fox News	Online	None of the above

Here we use add_stop_rule() to end the survey based on the response to an initial question (an option selection that makes the second question unnecessary):

from edsl import QuestionMultipleChoice, QuestionYesNo, Survey

# From the CES pre-election questionnaire
q_employ = QuestionMultipleChoice(
    question_name="employ",
    question_text="Which of the following best describes your current employment status?",
    question_options=[
        "Working full time now",
        "Working part time now",
        "Temporarily laid off",
        "Unemployed",
        "Retired",
        "Permanently disabled",
        "Taking care of home or family",
        "Student",
        "Other",
    ],
)

q_hadjob = QuestionYesNo(
    question_name="hadjob",
    question_text="At any time over the past five years, have you had a job?",
)

survey = Survey([q_employ, q_hadjob]).add_stop_rule(
    q_employ,
    "employ in ['Working full time now', 'Working part time now', 'Temporarily laid off']",
)

results = survey.by(agents).run()

results.select("employ", "hadjob")

	answer.employ	answer.hadjob
0	Working full time now	nan
1	Student	Yes
2	Working full time now	nan

Combining survey methods

Here we apply multiple methods at once: we add a memory of region to the prompt for inputzip, then pipe the answer to inputzip into the question text of votereg_f, and a stop rule if the answer to votereg is not “Yes” (i.e., do not administer votereg_f):

from edsl import QuestionMultipleChoice, QuestionList, QuestionYesNo, Survey

# From the CES pre-election questionnaire
q_region = QuestionMultipleChoice(
    question_name="region",
    question_text="In which census region do you live?",
    question_options=["Northeast", "Midwest", "South", "West"],
)

q_inputzip = QuestionList(
    question_name="inputzip",
    question_text="So that we can ask you about the news and events in your area, in what zip code do you currently reside?",
    max_list_items=1,
)

q_votereg = QuestionMultipleChoice(
    question_name="votereg",
    question_text="Are you registered to vote?",
    question_options=["Yes", "No", "Don't know"],
)

q_votereg_f = QuestionYesNo(
    question_name="votereg_f",
    question_text="Is {{ inputzip.answer[0] }} the zip code where you are registered to vote?",
)

survey = (
    Survey([q_region, q_inputzip, q_votereg, q_votereg_f])
    .add_targeted_memory(q_inputzip, q_region)
    .add_stop_rule(q_votereg, "votereg != 'Yes'")
)

results = survey.by(agents).run()

results.select("region", "inputzip", "votereg", "votereg_f_user_prompt", "votereg_f")

	answer.region	answer.inputzip	answer.votereg	prompt.votereg_f_user_prompt	answer.votereg_f
0	West	[94016]	Yes	Is 94016 the zip code where you are registered to vote? No Yes Only 1 option may be selected. Please respond with just your answer. After the answer, you can put a comment explaining your response.	No
1	South	[33101]	Yes	Is 33101 the zip code where you are registered to vote? No Yes Only 1 option may be selected. Please respond with just your answer. After the answer, you can put a comment explaining your response.	No
2	South	[30301]	Yes	Is 30301 the zip code where you are registered to vote? No Yes Only 1 option may be selected. Please respond with just your answer. After the answer, you can put a comment explaining your response.	No

See more details on all survey methods.

Parameterizing questions

We can create variations of questions using Scenario objects for content that we want to add to questions. This allows us to efficiently administer multiple versions of questions at once. We start by using a {{ parameter }} in a question:

from edsl import QuestionMultipleChoice, Survey

# Modified from the CES pre-election questionnaire
q_votereg = QuestionMultipleChoice(
    question_name="votereg",
    question_text="Are you {{ scenario.status }}?",
    question_options=["Yes", "No", "Don't know"],
)

survey = Survey([q_votereg])

Next we create a Scenario for each text that we want to insert in the question:

from edsl import Scenario, ScenarioList

statuses = [
    "registered to vote",  # original CES question
    "enrolled in school",
    "employed full- or part-time",
    "married or in a domestic partnership",
    "licensed to drive",
]

scenarios = ScenarioList(Scenario({"status": s}) for s in statuses)

results = survey.by(scenarios).by(agents).run()

(
    results.sort_by("status", "age")
    .select("age", "education", "sex", "status", "votereg")
)

	agent.age	agent.education	agent.sex	scenario.status	answer.votereg
0	18-29	High school or less	Male	employed full- or part-time	Yes
1	30-44	College/graduate	Female	employed full- or part-time	Yes
2	30-44	College/graduate	Female	employed full- or part-time	Yes
3	18-29	High school or less	Male	enrolled in school	Yes
4	30-44	College/graduate	Female	enrolled in school	No
5	30-44	College/graduate	Female	enrolled in school	No
6	18-29	High school or less	Male	licensed to drive	Don’t know
7	30-44	College/graduate	Female	licensed to drive	Don’t know
8	30-44	College/graduate	Female	licensed to drive	Yes
9	18-29	High school or less	Male	married or in a domestic partnership	No
10	30-44	College/graduate	Female	married or in a domestic partnership	No
11	30-44	College/graduate	Female	married or in a domestic partnership	Yes
12	18-29	High school or less	Male	registered to vote	Yes
13	30-44	College/graduate	Female	registered to vote	Yes
14	30-44	College/graduate	Female	registered to vote	Yes

Introduction

Getting Started

Core Concepts

Getting Data

Working with Results

Validating with Humans

No-code Apps

Coop

How-to Guides

Notebook Examples

Developers

Contents

Sample data: Cooperative Election Study Common Content, 2022

Reference & contact

Technical setup

Creating an AI agent

Designing agent panels

Agent instructions

Creating questions

Administering a survey

Answer commentary

Selecting language models

Question context & memory

Piping questions

Survey rules & conditions

Combining survey methods

Parameterizing questions

Introduction

Getting Started

Core Concepts

Getting Data

Working with Results

Validating with Humans

No-code Apps

Coop

How-to Guides

Notebook Examples

Developers

​Contents

​Sample data: Cooperative Election Study Common Content, 2022

​Reference & contact

​Technical setup

​Creating an AI agent

​Designing agent panels

​Agent instructions

​Creating questions

​Administering a survey

​Answer commentary

​Selecting language models

​Question context & memory

​Piping questions

​Survey rules & conditions

​Combining survey methods

​Parameterizing questions

Contents

Sample data: Cooperative Election Study Common Content, 2022

Reference & contact

Technical setup

Creating an AI agent

Designing agent panels

Agent instructions

Creating questions

Administering a survey

Answer commentary

Selecting language models

Question context & memory

Piping questions

Survey rules & conditions

Combining survey methods

Parameterizing questions