Aria - Headless AI Agents

Aria - our Teacher’s application processing multi-agent systems involve multiple autonomous agents that interact and collaborate with each other to perform complex tasks and work behind the scenes.

Comment 1

Inspiration

We wanted to explore the Agentforce platform and the capabilities it supports by creating a real-life project. This Hackathon gave us that opportunity to go into the depth of the platform and work on a hands-on development project. We chose to use RAG features of Data Cloud and Headless multi-agent architecture to build this solution.

What it does

Aria as a headless agent does the following tasks:

  • Picks the submitted teacher's application on a scheduled basis and process one by one without any user intervention
  • Main agent: orchestrates various tasks performed by different autonomous agents
  • Application screening Agent: Do initial screen of the applications submitted, verify the field values etc.
  • Document Validation Agent: Validate the documents attached for the Application
  • Application decision Agent: Update the applications’ status
  • Case creation Agent: Create a case for the application

How we built it

  • We have used Agent Builder to create service agent to perform a specific set of tasks for each agent as mentioned above
  • A scheduled Flow to trigger the main agent and autolaunced flows to invoke the sub-agents
  • A custom integration solution is built to store SF attachment files into AWS S3 and then ingesting unstructured data into Data Cloud
  • A RAG solution is built by integrating AWS S3 bucket and Data Cloud to validate application attachment documents
  • A custom retriever is written to retrieve contents specific to an attachment e.g. we had to query in qualification doc that is attached for application 1 not with application 2
  • An agent orchetration logic is built to call sub-agents to perform specific tasks to automate the whole application process

Challenges we ran into

  • Ingesting SF attachment into Data Cloud was not working in this Org, had to try setting up in another developer org but even though we made it work after raising a SF case (#470180073), the current feature has limitation that if we need to query on a specific file there is no way we can differentiate which vector chunk is associated with which attachment file. This took a lot of time to figure out the issue that we ran into to create the search index and then how to retrieve it. Finally, we had to implement the alternate approach with S3 ingestion.
  • Instructions given in Create an Unstructured Data Connection from Amazon S3 to setup Amazon S3 file notification also took a lot of time, had to fix the installer script issue as there are not much resources available.
  • Agent to agent communication is failing as the Service Agent user can't access another agent due to perm issue, we still have an open case (#470370742), a product bug is causing this issue. As a workaround we are running the container flow in System mode to make the solution working which we need to re-test after the bug fix is in place.
  • We have tried severals options to see how the Main orchestration agent can communicate with other agents and chain the entire processing logic, that took some time to understand how the Reasoning engine and the agentic framework work.

Accomplishments that we're proud of

  • As we came to know about this Hackathon at the end of March, within a few days, having only 2 members in the team, we were able to complete 5 Agents, 7 flows, and 1 apex
  • Salesforce, Amazon S3 and Data Cloud integration to make the RAG solution working for our use cases

What we learned

  • We learnt how to approach an AI project and how important the business benefits are for the use cases.
  • How important the Topic and agent instructions are
  • RAG Hybrid and Vector search capabilities of Data Cloud
  • S3 ingestion into data Cloud and how to build a custom retreiver to overcome some of the limitations of default retriever

What's next for Aria - Headless AI Agents

  • We will automate some more verifications, like police verification etc. by implementing the third-party integration (example: Police verification system)
  • Agent API integration to communicate between Agentforce Agent with third-party agents.
  • We would like to use this prototype solution to demonstrate to our prospective clients working in Financial, Insurance, Telecommunication, Public sector domain to see what all opportunities we have to implement agentforce solutions for them.

Built With

Share this project:

Updates