project-dunder-mifflin

Chat with Boss Agent
Chat in conference Room with all agents in
one to one chat where the agent communicates with remote agent via A2A protocol on the fly

Inspiration

The idea of leveragfing AI into supportind our day to day work, the raise of agents inspired me to work on this sort of project but i didn't find enough motivation to start working on this until i saw the hackathon announcement, so i took that as an inspiration to set a deadline to learn AI agents systems and put my ideas into work as well. and the idea to incorporate the office into this inspired me heavily

What it does

This is a replica of real world IT Team there's a developer agent, a tester agent, an HR agent, a support agent and a security agent and also a data science agent working together to support the real people working in IT teams.

There are 3 scenarios we could use this app

One to one conversation

We can engage in conversation with any single agent for example ask the developer agent to work on a feature while also asking the database agent to generate reports from database, just like we chat with different people in office for differnt works

Conversation with the Boss

In this project Michael Scarn acts as the boss, who has all the other agents as tool so he can work with all of them and perform what we wanted, for example, we can provide him a detailed task of involving all the agents, he breaks down them and use each agent to get it done, he can also communicate back to the agents and get us the task ready finally

Conference Room

This is more of a meeting room chat, where we can summon any agent we want and other agents can talk wigth different agents in the group, here all agents will work as a sub agentso that we can talk with anyone we want, for example we can ask holly to look for new recruits and once she privides the contact info we can ask michael to check with the new recruit about his information, michael can use the url and initiate A2A protocol for conversation and does the job

How we built it

I built it with ADK mainly, and deployed it in Cloud run, The original plan was more similar but then it evolved along the way and now it was at this stage.

Challenges we ran into

I had issues trying to orchestrate the multi agent communication and delegation tasks back and forth but figured things with multiple drafts of prompts and different orchestration methods like agent as a tool wherever required

Accomplishments that we're proud of

I'm proud of this project entirely, as i was able to mimic an entire IT team with ai agents in a way they can support real world people and get things done quickly. But incoporating The Office characters was one thing i was more happy about

What we learned

If thought properly we can design anything with AI by breaking it into smaller chunk, like with this project each agent and it's sub agents are working on smaller tasks while acheiving a big set of goals,