DEV Community

Miles Georgi
Miles Georgi

Posted on

Writing an AI Agent in 1 Line of Ruby Code Using Foobara's AgentBackedCommand

Writing an AI Agent in 1 Line of Ruby Code Using Foobara's AgentBackedCommand

In this article, we'll introduce/demo Foobara::AgentBackedCommand by adding it to a program with an existing Foobara domain in it. Let's do it!

What is Foobara?

Foobara is a command-centric and discoverable software framework that helps manage domain complexity by abstracting away integration code and other features.

The Demo Loan Origination Domain and Demo Records

First, we need a program to add it to. Here is a program that uses a demo loan origination domain written using Foobara. You can find the demo domain implementation here: https://github.com/foobara-demo/loan-origination/tree/main/src

Here's our program in an executable file called loan-origination:

#!/usr/bin/env ruby

require_relative "boot"
require "foobara/sh_cli_connector"

connector = Foobara::CommandConnectors::ShCliConnector.new
connector.connect(FoobaraDemo::LoanOrigination)
connector.run
Enter fullscreen mode Exit fullscreen mode

This uses a Foobara CLI command connector so that we can run commands from LoanOrigination on the command line.

Let's see what commands are available:

$ ./loan-origination 
Usage: loan-origination [GLOBAL_OPTIONS] [ACTION] [COMMAND_OR_TYPE] [COMMAND_INPUTS]
...
Available commands:
  FoobaraDemo::LoanOrigination::SubmitApplicationForUnderwriterReview
  FoobaraDemo::LoanOrigination::FindALoanFileThatNeedsReview            Will return a loan file
                                                                        that needs review or nil
                                                                        if there are no loan 
                                                                        files that need review
  FoobaraDemo::LoanOrigination::FindAllLoanFiles
  FoobaraDemo::LoanOrigination::GenerateLoanFilesReport
  FoobaraDemo::LoanOrigination::StartUnderwriterReview
  FoobaraDemo::LoanOrigination::StartLoanApplication
  FoobaraDemo::LoanOrigination::CreateCreditPolicy
  FoobaraDemo::LoanOrigination::FindCreditPolicy
  FoobaraDemo::LoanOrigination::ApproveLoanFile
  FoobaraDemo::LoanOrigination::AddCreditScore
  FoobaraDemo::LoanOrigination::FindLoanFile
  FoobaraDemo::LoanOrigination::DenyLoanFile
  FoobaraDemo::LoanOrigination::AddPayStub                              Add a pay stub to a loan
                                                                        file
  FoobaraDemo::LoanOrigination::Demo::PrepareDemoRecords
$
Enter fullscreen mode Exit fullscreen mode

We can see a ton of commands at our disposal. Let's run GenerateLoanFilesReport:

$ ./loan-origination GenerateLoanFilesReport
$
Enter fullscreen mode Exit fullscreen mode

We get nothing back. Which makes sense! Let's create some records to play with using PrepareDemoRecords:

$ ./loan-origination PrepareDemoRecords
...
$
Enter fullscreen mode Exit fullscreen mode

This prints out a bunch of loan file data that is excluded. Now let's look at our report again:

$ ./loan-origination GenerateLoanFilesReport
{
  id: 1,
  applicant: "Barbara",
  state: "needs_review"
},
{
  id: 2,
  applicant: "Basil",
  state: "needs_review"
},
{
  id: 3,
  applicant: "Fumiko",
  state: "needs_review"
}
$
Enter fullscreen mode Exit fullscreen mode

This command creates 3 loan files ready for review that are all on different credit policies. We expect these loan files to be denied or approved for different reasons. See PrepareDemoRecords for details on the demo records and the expected results of reviewing these loan files.

Now, we could review these loan files ourselves using loan-origination since it exposes all the commands that we need. We could do the following:

$ ./loan-origination FindALoanFileThatNeedsReview
...
id: 1
$
Enter fullscreen mode Exit fullscreen mode

We could then do ./loan-origination StartUnderwriterReview --loan-file 1 which would start our review.

We would then need to do ./loan-origination FindCreditPolicy --credit-policy 1 and finally we could run ./loan-origination DenyLoanFile --loan-file 1 --credit-score-used 650 --denied-reasons low_credit_score.

We could then move on to the next loan file until no more need review.

In this case, we're a human operating the domain, which we've chosen in this case to expose through a CLI.

If we had an AI agent that also knew how to operate our domain, we could say something like: "Hi! Could you please review every loan file that needs review? Thanks!"

Notice that what we would be doing is issuing a high-level domain operation to our AI agent. Since Foobara commands are meant to encapsulate high-level domain operations, it would be nice if we could just have a ReviewAllLoanFiles Foobara command but without having to write its #execute method (ie, the domain logic itself.)

Defining and Running a Foobara::AgentBackedCommand

Now it's finally time to write our AI agent in 1 line of Ruby code, introducing Foobara::AgentBackedCommand:

#!/usr/bin/env ruby

require_relative "boot"
require "foobara/sh_cli_connector"

class FoobaraDemo::LoanOrigination::ReviewAllLoanFiles < Foobara::AgentBackedCommand; end

connector = Foobara::CommandConnectors::ShCliConnector.new
connector.connect(FoobaraDemo::LoanOrigination)
connector.run
Enter fullscreen mode Exit fullscreen mode

And we're done! We've created a command called ReviewAllLoanFiles which is an AgentBackedCommand.

We can now see it in our loan-origination --help output:

$ ./loan-origination --help
...
Available commands:

  FoobaraDemo::LoanOrigination::SubmitApplicationForUnderwriterReview
  FoobaraDemo::LoanOrigination::FindALoanFileThatNeedsReview            Will return a loan file
                                                                        that needs review or nil if
                                                                        there are no loan files
                                                                        that need review
  FoobaraDemo::LoanOrigination::FindAllLoanFiles
  FoobaraDemo::LoanOrigination::GenerateLoanFilesReport
  FoobaraDemo::LoanOrigination::StartUnderwriterReview
  FoobaraDemo::LoanOrigination::StartLoanApplication
  FoobaraDemo::LoanOrigination::CreateCreditPolicy
  FoobaraDemo::LoanOrigination::FindCreditPolicy
  FoobaraDemo::LoanOrigination::ApproveLoanFile
  FoobaraDemo::LoanOrigination::AddCreditScore
  FoobaraDemo::LoanOrigination::FindLoanFile
  FoobaraDemo::LoanOrigination::DenyLoanFile
  FoobaraDemo::LoanOrigination::AddPayStub                              Add a pay stub to a loan
                                                                        file
  FoobaraDemo::LoanOrigination::Demo::PrepareDemoRecords
  FoobaraDemo::LoanOrigination::ReviewAllLoanFiles
...
$
Enter fullscreen mode Exit fullscreen mode

We can see it there at the bottom. Let's see what inputs it has:

$ ./loan-origination --help ReviewAllLoanFiles
...
Command inputs:

 -v, --agent-options-verbose
...
 -a, --agent-name AGENT_NAME
 -l, --llm-model LLM_MODEL                             The model to use for the
  LLM. One of: chatgpt-4o-latest, claude-2.0, claude-2.1, 
  ...
  claude-3-5-sonnet-20241022, claude-3-7-sonnet-20250219, 
  ...
  claude-opus-4-20250514, claude-sonnet-4-20250514, gpt-3.5-turbo,
  ... 
  gpt-4o, gpt-4o-2024-05-13, gpt-4o-2024-08-06, gpt-4o-2024-11-20,
  ...
  Default: claude-3-7-sonnet-20250219
...
$
Enter fullscreen mode Exit fullscreen mode

I've snipped many options as well as a ton of models out of the --llm-model option. I won't make use of the --llm-model option here but it supports ollama, openai, and anthropic models. It defaults to claude-3-7-sonnet-20250219.

I will make use of --agent-options-verbose so we can see what decisions it makes, as well as --agent-name, though. So... let's review these loan files!

$ ./loan-origination ReviewAllLoanFiles --agent-options-verbose --agent-name UnderwritingAgent
UnderwritingAgent: Foobara::Agent::DescribeCommand.run(command_name: "Foobara::Agent::ListCommands")
UnderwritingAgent: Foobara::Agent::ListCommands.run
UnderwritingAgent: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::FindAllLoanFiles")
UnderwritingAgent: FoobaraDemo::LoanOrigination::FindAllLoanFiles.run
UnderwritingAgent: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::StartUnderwriterReview")
UnderwritingAgent: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::FindCreditPolicy")
UnderwritingAgent: FoobaraDemo::LoanOrigination::StartUnderwriterReview.run(loan_file: 1)
UnderwritingAgent: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::FindLoanFile")
UnderwritingAgent: FoobaraDemo::LoanOrigination::FindLoanFile.run(loan_file: 1)
UnderwritingAgent: FoobaraDemo::LoanOrigination::FindCreditPolicy.run(credit_policy: 1)
UnderwritingAgent: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::ApproveLoanFile")
UnderwritingAgent: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::DenyLoanFile")
UnderwritingAgent: FoobaraDemo::LoanOrigination::DenyLoanFile.run(loan_file: 1, credit_score_used: 650, denied_reasons: ["low_credit_score"])
UnderwritingAgent: FoobaraDemo::LoanOrigination::StartUnderwriterReview.run(loan_file: 2)
UnderwritingAgent: FoobaraDemo::LoanOrigination::FindLoanFile.run(loan_file: 2)
UnderwritingAgent: FoobaraDemo::LoanOrigination::FindCreditPolicy.run(credit_policy: 2)
UnderwritingAgent: FoobaraDemo::LoanOrigination::DenyLoanFile.run(loan_file: 2, credit_score_used: 750, denied_reasons: ["insufficient_pay_stubs_provided"])
UnderwritingAgent: FoobaraDemo::LoanOrigination::StartUnderwriterReview.run(loan_file: 3)
UnderwritingAgent: FoobaraDemo::LoanOrigination::FindLoanFile.run(loan_file: 3)
UnderwritingAgent: FoobaraDemo::LoanOrigination::FindCreditPolicy.run(credit_policy: 3)
UnderwritingAgent: FoobaraDemo::LoanOrigination::ApproveLoanFile.run(loan_file: 3, credit_score_used: 750)
UnderwritingAgent: FoobaraDemo::LoanOrigination::FindAllLoanFiles.run
UnderwritingAgent: Foobara::Agent::NotifyUserThatCurrentGoalHasBeenAccomplished.run
$
Enter fullscreen mode Exit fullscreen mode

Here we can see all the decisions it made. Let's look at the report now:

$ ./loan-origination GenerateLoanFilesReport
{
  id: 1,
  applicant: "Barbara",
  state: "denied",
  underwriter_decision: {
    decision: :denied,
    credit_score_used: 650,
    denied_reasons: [
      :low_credit_score
    ]
  }
},
{
  id: 2,
  applicant: "Basil",
  state: "denied",
  underwriter_decision: {
    decision: :denied,
    credit_score_used: 750,
    denied_reasons: [
      :insufficient_pay_stubs_provided
    ]
  }
},
{
  id: 3,
  applicant: "Fumiko",
  state: "drafting_docs",
  underwriter_decision: {
    decision: :approved,
    credit_score_used: 750
  }
}
$
Enter fullscreen mode Exit fullscreen mode

Looks great! It used the correct credit scores for the different loan files, made the correct approved/denied decisions and chose the proper denied reason for the different denied loan files.

A More Complex and Interesting Example

Adding an agent to a program with one line of Ruby code might be fun clickbait and all, but let's look at an example that highlights some features of AgentBackedCommand that are not demonstrated in the one-liner example above. Check this out, in a new executable called review-all-loan-files we have written:

#!/usr/bin/env ruby

require_relative "boot"

module FoobaraDemo
  module LoanOrigination
    class UnderwriterSummary < Foobara::Model
      attributes do
        loan_file_id :integer, :required
        pay_stub_count :integer, :required
        fico_scores [:integer, :integer, :integer], :required
        credit_policy CreditPolicy, :required
      end
    end

    class ReviewLoanFile < Foobara::AgentBackedCommand
      description "Starts the underwriter review then checks requirements in its CreditPolicy " \
                  "and approves or denies accordingly."

      inputs UnderwriterSummary
      result LoanFile::UnderwriterDecision

      depends_on StartUnderwriterReview, DenyLoanFile, ApproveLoanFile

      verbose
      llm_model "claude-3-7-sonnet-20250219"
      agent_name "Inner"
    end

    class ReviewAllLoanFiles < Foobara::AgentBackedCommand
      result [{
                applicant_name: :string,
                decision: LoanFile::UnderwriterDecision
              }]

      depends_on ReviewLoanFile, FindALoanFileThatNeedsReview

      verbose
      llm_model "gpt-4o"
      agent_name "Outer"
    end
  end
end

outcome = FoobaraDemo::LoanOrigination::ReviewAllLoanFiles.run

if outcome.success?
  outcome.result.each do |name_and_decision|
    name, decision = name_and_decision.values_at(:applicant_name, :decision)

    if decision.denied?
      puts "Denied: #{name}"
      puts "Reason: #{decision.denied_reasons.join(", ")}"
    else
      puts "Approved: #{name}"
    end
    puts
  end
else
  puts outcome.errors_hash
end
Enter fullscreen mode Exit fullscreen mode

We've abandoned the CLI command connector and are just using the commands directly.

We have introduced a second AgentBackedCommand called ReviewLoanFile. These two AgentBackedCommands are not only using two different models but two different services entirely.

Specifying an Inputs Type for an AgentBackedCommand

We are making use of inputs on ReviewLoanFile just like we would with any Foobara::Command that needs inputs. Originally, inputs there were defined as:

inputs do
  loan_file LoanFile, :required
end
Enter fullscreen mode Exit fullscreen mode

This worked but this presents a problem. Can you spot it? We want to avoid things like bias from the names of the applicants influencing decisions. So we remodeled it by introducing a new model to provide the minimum information required and called it UnderwriterSummary.

Something cool about this is we didn't have to change ReviewAllLoanFiles when we changed the inputs to ReviewLoanFile since it has no #execute method nor domain logic to update to conform to the new interface!

Specifying and Using the AgentBackedCommand Result Programmatically!

We are also specifying a result type for ReviewAllLoanFiles, like so:

result [{
          applicant_name: :string,
          decision: LoanFile::UnderwriterDecision
        }]
Enter fullscreen mode Exit fullscreen mode

and are making programmatic use of its outcome like so:

if decision.denied?
  puts "Denied: #{name}"
  puts "Reason: #{decision.denied_reasons.join(", ")}"
else
  puts "Approved: #{name}"
end
Enter fullscreen mode Exit fullscreen mode

We are guaranteed that this data has the types we specified in our result declaration, just like with any other Foobara::Command and can use it programmatically just fine as a result.

We are also making use of depends_on which is standard Foobara stuff to limit what subcommands a command is allowed to run.

There are a few AgentBackedCommand-specific DSL methods that we're making use of here: verbose will let us see which decisions which Agent is making, llm_model will let us specify the model for each agent independently, and agent_name will let us specify the name for each agent independently so we can tell them apart in the output.

Let's first re-create our demo records:

$ ./loan-origination Demo::PrepareDemoRecords
Enter fullscreen mode Exit fullscreen mode

And now let's run this new script and review these loan files with two agents!

$ ./review-all-loan-files
Outer: Foobara::Agent::DescribeCommand.run(command_name: "Foobara::Agent::ListCommands")
Outer: Foobara::Agent::ListCommands.run
Outer: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::FindALoanFileThatNeedsReview")
Outer: FoobaraDemo::LoanOrigination::FindALoanFileThatNeedsReview.run
Outer: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::ReviewLoanFile")
Outer: FoobaraDemo::LoanOrigination::ReviewLoanFile.run(loan_file_id: 4, pay_stub_count: 1, fico_scores: [600, 650, 750], credit_policy: 4)
Inner: Foobara::Agent::DescribeCommand.run(command_name: "Foobara::Agent::ListCommands")
Inner: Foobara::Agent::ListCommands.run
Inner: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::StartUnderwriterReview")
Inner: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::DenyLoanFile")
Inner: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::ApproveLoanFile")
Inner: FoobaraDemo::LoanOrigination::StartUnderwriterReview.run(loan_file: 4)
Inner: FoobaraDemo::LoanOrigination::DenyLoanFile.run(loan_file: 4, credit_score_used: 650, denied_reasons: ["low_credit_score"])
Inner: Foobara::Agent::DescribeCommand.run(command_name: "Foobara::Agent::Inner::NotifyUserThatCurrentGoalHasBeenAccomplished")
Inner: Foobara::Agent::Inner::NotifyUserThatCurrentGoalHasBeenAccomplished.run(result: {"decision" => "denied", "credit_score_used" => 650, "denied_reasons" => ["low_credit_score"]})
Outer: FoobaraDemo::LoanOrigination::FindALoanFileThatNeedsReview.run
Outer: FoobaraDemo::LoanOrigination::ReviewLoanFile.run(loan_file_id: 5, pay_stub_count: 1, fico_scores: [600, 650, 750], credit_policy: 5)
Inner: Foobara::Agent::DescribeCommand.run(command_name: "Foobara::Agent::ListCommands")
Inner: Foobara::Agent::ListCommands.run
Inner: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::StartUnderwriterReview")
Inner: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::DenyLoanFile")
Inner: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::ApproveLoanFile")
Inner: FoobaraDemo::LoanOrigination::StartUnderwriterReview.run(loan_file: 5)
Inner: FoobaraDemo::LoanOrigination::DenyLoanFile.run(loan_file: 5, credit_score_used: 750, denied_reasons: ["insufficient_pay_stubs_provided"])
Inner: Foobara::Agent::DescribeCommand.run(command_name: "Foobara::Agent::Inner::NotifyUserThatCurrentGoalHasBeenAccomplished")
Inner: Foobara::Agent::Inner::NotifyUserThatCurrentGoalHasBeenAccomplished.run(result: {"decision" => "denied", "credit_score_used" => 750, "denied_reasons" => ["insufficient_pay_stubs_provided"]})
Outer: FoobaraDemo::LoanOrigination::FindALoanFileThatNeedsReview.run
Outer: FoobaraDemo::LoanOrigination::ReviewLoanFile.run(loan_file_id: 5, pay_stub_count: 1, fico_scores: [600, 650, 750], credit_policy: 5)
Inner: Foobara::Agent::DescribeCommand.run(command_name: "Foobara::Agent::ListCommands")
Inner: Foobara::Agent::ListCommands.run
Inner: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::StartUnderwriterReview")
Inner: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::DenyLoanFile")
Inner: Foobara::Agent::DescribeCommand.run(command_name: "FoobaraDemo::LoanOrigination::ApproveLoanFile")
Inner: FoobaraDemo::LoanOrigination::StartUnderwriterReview.run(loan_file: 5)
Inner: FoobaraDemo::LoanOrigination::ApproveLoanFile.run(loan_file: 5, credit_score_used: 750)
Inner: Foobara::Agent::DescribeCommand.run(command_name: "Foobara::Agent::Inner::NotifyUserThatCurrentGoalHasBeenAccomplished")
Inner: Foobara::Agent::Inner::NotifyUserThatCurrentGoalHasBeenAccomplished.run(result: {"decision" => "approved", "credit_score_used" => 750})
Outer: FoobaraDemo::LoanOrigination::FindALoanFileThatNeedsReview.run
Outer: Foobara::Agent::DescribeCommand.run(command_name: "Foobara::Agent::Outer::NotifyUserThatCurrentGoalHasBeenAccomplished")
Outer: Foobara::Agent::Outer::NotifyUserThatCurrentGoalHasBeenAccomplished.run(result: [{"applicant_name" => "Barbara", "decision" => {"decision" => "denied", "credit_score_used" => 650, "denied_reasons" => ["low_credit_score"]}}, {"applicant_name" => "Basil", "decision" => {"decision" => "denied", "credit_score_used" => 750, "denied_reasons" => ["insufficient_pay_stubs_provided"]}}, {"applicant_name" => "Fumiko", "decision" => {"decision" => "approved", "credit_score_used" => 750}}])

Denied: Barbara
Reason: low_credit_score

Denied: Basil
Reason: insufficient_pay_stubs_provided

Approved: Fumiko
Enter fullscreen mode Exit fullscreen mode

Cool! We programmatically used result data from a command without having to write its execute method!

And if we check ./loan-origination GenerateLoanFilesReport we can see that our program did everything correctly!

An Interesting Takeaway

Putting together this demo, something crossed my mind: It could be that we might find ourselves wanting to use LLMs to handle some of this domain logic for us as we prototype and discover the domain, but once the business scales, it might be worth the cost savings to then convert AgentBackedCommands to just regular Foobara::Commands.

Take for example ReviewAllLoanFiles. We did get some benefit of not having to manage the interface change between it and ReviewLoanFile`, but it could be expensive to run if the business scales even though its domain logic isn't that difficult to implement in a typical #execute method.

This possibility is interesting to me because it is the opposite of my intuition: that we might start with a well-working domain and then look for opportunities to automate parts of it. In some cases, it might be that we want the reverse: to automate various high-level domain operations during prototyping and early stages of business and then start replacing LLM automation with implemented domain logic as the business scales and the domain solidifies.

Links

Please Reach Out!

Want help using Foobara or want to help Foobara by contributing code/art/documentation/whatever to it? Please reach out! https://foobara.com

Disclaimers

1) This agent just does whatever it thinks it needs to do to accomplish its goal. So play with this with caution!
2) This loan origination domain is just to demo cool aspects of Foobara in a relatable way. This is probably obvious, but this is not meant for making real credit decisions, which might not even be a safe task for an LLM, anyway.

Top comments (0)