Agent Bond | Devpost

Inspiration

Agent Bond was inspired by the experience of attending a large event like Web Summit, where thousands of people from different backgrounds come together with one common goal: to make meaningful connections.

At an event with more than 20,000 attendees, finding the right people can feel overwhelming. Even with filters like industry or role, it is still hard to pinpoint people with similar backgrounds, shared interests, or complementary goals.

We wanted to solve that problem by creating an agent that helps attendees connect with the people who matter most to them.

What it does

Agent Bond helps event attendees find relevant people to connect with faster.

Users create a lightweight digital twin that captures who they are, what they are looking for, what they can offer, their interests, skills, preferred conversation topics, and availability.

The agent then recommends potential connections, explains why each match is relevant, and suggests a conversation starter to make networking easier and more intentional.

Agent Bond also includes calendar integration, so schedules can be used as part of the matching criteria. This helps the agent recommend not only who someone should meet, but also when a connection may be easier to make.

How we built it

We started by creating a simple Functional Requirements Document (FRD), to define the scope of the project and keep the MVP realistic for a 3-day hackathon.

Then we used coding agents to move quickly through the build process.

The process included:

Defining the user flows
Designing the networking experience
Creating the app specification
Building the core matching and recommendation logic
Creating agent-to-agent communication
Integrating calendar availability into the matching process

For the matching logic, each person’s profile is converted into an embedding, which represents the profile as a multidimensional vector. The database then helps us find likely candidates using multidimensional geometry, based on the matching logic we defined.

The current match score is based on:

Shared interests: 30%
Goal alignment: 30%
Skills match: 20%
Industry relevance: 10%
Conversation topic overlap: 10%

We also explored sentiment analysis to help the agent better understand context, tone, and networking intent.

Challenges we ran into

One of the biggest challenges was creating the interface for agent-to-agent communication.

We had to think through how one attendee’s agent could communicate with another attendee’s agent in a way that felt useful, respectful, and relevant.

Another challenge was designing the agent experience itself. We needed to figure out how to make conversations feel natural while still keeping them focused on networking.

We also had to think carefully about matching logic. Right now, the matching is useful, but still a bit deterministic. The agent follows the criteria we defined, but we want future versions to bring in more context so recommendations feel more nuanced and human.

Accomplishments that we're proud of

We are proud that we were able to turn a common event problem into a simple, useful MVP.

In a short amount of time, we defined the product scope, created the user flows, designed the experience, built the matching logic, created agent-to-agent communication, and added calendar integration.

We are also proud that the recommendation logic is explainable. Agent Bond does not just suggest people randomly. It can show why a match makes sense based on shared interests, goal alignment, skills, industry relevance, conversation topics, and availability.

What we learned

We learned that networking is not just about matching people by job title or industry.

A meaningful connection depends on context:

what someone is looking for
what they can offer
what they are curious about
what they are available for
why a conversation would be valuable

We also learned that designing an agent experience is different from designing a regular app experience. The agent needs to understand intent, preserve context, and make recommendations that feel relevant, not random.

Another important learning was that embeddings are powerful, but the logic around them matters. A multidimensional profile match can help surface likely candidates, but the quality of the recommendation depends on how well we define the matching criteria and how much useful context the agent has.

What's next for Agent Bond

Next, we want to make Agent Bond smarter by giving the algorithm more context, so matches feel less deterministic and more human.

This could include giving users the option to enrich their profile with public profiles and interests from places like YouTube, Spotify, Reddit groups, and social media, while allowing each person to choose exactly what they want to share and how much context they want to provide.

We also want to improve sentiment analysis and agent-to-agent communication so recommendations feel more natural, timely, and personalized.

The long-term vision is for Agent Bond to become a personal networking companion that helps people connect before, during, and after events.