Touch grass, and touch glass on a padel court

#devchallenge #hf26challenge #mlhacks #tpf

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

Arranging a padel match can be surprisingly tedious. You need four players of roughly similar ability, available at the same time, preferably with compatible expectations. In practice, this means checking club apps, messaging WhatsApp groups, negotiating times, and chasing people for confirmation.

I wondered whether personal AI agents could handle some of that coordination for us.

PadelPot is an experiment where four AI agents negotiate a doubles match on behalf of their owners. Each agent knows its player's preferences, can discuss possible arrangements with other agents, and asks its owner when a compromise or commitment requires approval.

The agents don't exchange complete player profiles. They negotiate using limited information, with privacy and commitment rules enforced by the application rather than left to the language model.

The idea behind Touch Grass is straightforward: less time arranging matches, more time playing them.

Demo

The demo follows four fictional players—Alba, Nico, Luz, and Teo—using a WhatsApp-style chat interface. Their agents negotiate a Thursday evening match, resolve a disagreement about the starting time, and obtain approval from all four players.

The conversations use actual Gemma inference rather than scripted responses.

Code

https://github.com/The-Pipeline-Framework/padelpot

The application is open source and includes setup instructions, synthetic player profiles, automated tests, and reproducible demonstration scenarios.

How I Built It

PadelPot is a Java application built using The Pipeline Framework (TPF), an open-source framework I've been developing for distributed applications.

It uses Google's open-weight Gemma models for interviewing players and negotiating between agents. I tested the application using Gemma through Ollama, running locally. It also supports inference through OpenRouter.

TPF provides the typed pipelines, agentic execution, persistence, and asynchronous human interactions. The application handles the matchmaking rules, privacy policies, and final approvals.

One interesting test sends 30 concurrent approval attempts to the same match. Despite the competing requests, the application never confirms more than four players.

The first working implementation took a little over an hour using an AI-assisted development workflow, although it builds on TPF infrastructure developed over several years.

This is a working demonstration with synthetic users, not a production matchmaking service. Real WhatsApp integration, court bookings, and independent hosting of personal agents remain outside its scope.

Why Does Open Innovation Matter?

The interesting part of PadelPot isn't simply generating text with an LLM. It's allowing several independently acting agents to negotiate while keeping their behavior constrained by ordinary, testable application rules.

Open-weight models such as Gemma make it possible to experiment with that behavior locally, inspect the results, change inference providers, and run integration tests without depending exclusively on a proprietary model API.

TPF is also open source, so the execution model and application architecture can be inspected, tested, and extended.

For an experiment involving personal information and autonomous agents, that openness matters.