Barka

Project Managers should be able to define what's being done, who's doing it and how far along they are with enough visibility to adjust and stay on track, - Minus the Hustle. That's Barka

Comment

Inspiration

As a software developer, I have worked in multiple software and design agencies. And almost always, there was this rush hour thing for projects, client frustrations just to mention a few, but the most pressing issue was how I as a dev was being micromanaged and overloaded with tasks and ending up to receive less than or equal compensation for the work done.

We would have this meetings everyday which took up upto 2 hours or more discussing projects and tasks that are mostly overdue. The agencies and dev shops mostly used custom PM software that relied on static data and had this learning curve and no automated tracking of work and tasks across teams. Scope creep was innevitable.

With AI agents, this should be a breeze and a thing of the past.

What it does

Not Asana, Not Clickup, Not Notion. Barka is a platform that has been built to scale with AI agents as the operating system. We have found a way to build AI agents into software agents that offers enough agency to PM's, devs and clients and manages communication, tasks and projects across multiple stakeholders.

We started with building a collection of 5 agents: Gaia - the orchestrator agent that routes tasks to other sub agents

Documentation Agent - Handles all document generation tasks, like SRS documents, SOW projects, project proposals, contracts and will handle more in the future Discovery Agent - Handles client communication and onboarding of a client for a project and collects all client info, no need for a human to have endless meetings to understand clients needs and preferences Scheduling Agent - Handles meetings that follow organisation policies in setting up meetings usig google calendar and syncing them with the database

(And now, the cream) Project Manager Agent - This is the Junior Chief of Staff. Handles tasks creation, project management, progress tracking, team member tracking and more. This agent has access to a custom MCP server that gives it the context it needs like a human PM would about projects in an organisation.

How we built it

Barka has been built using a nodejs backend which acts as a proxy between the frontend and ovara agents ( the group of agents )

Challenges we ran into

1. Custom MCP

Given that google-adk is a new technology, some of the concepts like building custom MCP servers was really hectic. We had to implement a custom patch to increase the pre-set timeout for MCP

2. State Injection and agent callbacks

Being able to instruct an agent into using the correct state from session, especially with the custom MCP server was a major roadblock. This took a lot of research and reworking of the codebase to inject user info like id's roles etc which had to be done via the backend proxy.

3. Hosting

We had to convert our whole codebase from a polyrepo to a monorepo (something we plan to reconsider) in order to have a streamlined deployment pipeline so as to host on our server.

Trivial but important

One of the major regrets we have was building our whole codebase using MongoDB. We realised too late into the project that google-adk supports relational databases out of the box. So building a custom DatabaseSessionService came with challenges and setbacks

Accomplishments that we're proud of

1. Agents-UI sync

Building a custom event driven architecture that allows the agents to manipulate real-world data about projects, tasks and teams was something that we couldnt help but be overly proud of. For instance, the PM might want to arrange meetings with the team members with backlogs without the hustle of communicating with all of them, so making the project manager agent lias with the scheduling agent and passing context between them and immediately seeing calendar and UI updates was truly marvel for us

2. Beautifully crafted UI

The overall UI for Barka is something that really outshines most PM software. It allows visibility and gives true agency to teams and PM's in software companies by making it easy to switch between the agent interface and the traditional interface in case the text is too much.

3. Prompt engineering

We watched the project manager's prompt grow from 50 lines to 1134 lines and back to approx 199 lines after finetuning and doing some evals for the agents. Even though the agent still has some issues in interpetting what tool call should be done, it has really evolved.

What we learned

Comparing google-adk and other Agent building platforms, I think google-adk offers the best and simplest, most straightforward way to building something close to a true Multi-Agent-System.

When building a MAS, next time, We prefer building using a relational database for the purposes of consistent data fetching for context processing for agents

What's next for Barka

Barka is poised to build a sotware agencies focused PM software. We plan to intergrate MCP tools for memory, notion, github and in a brighter future, have partneships with companies building remote agents so that to allow true autonomy.

The dream is to allow small dev shops, software companies and indie developers to operate at the scale of thousands by offering truly autonomous AI agents as employees that dilligently work for holistic goal to serve their purposes, in a cost effective way and most importantly... MINUS THE HUSTLE

Built With

Share this project:

Updates