So you think you can Agent?
Highlights and lessons from the first Fixie Hackathon in Seattle
On Saturday, March 11th we hosted more than 50 Large Language Models(LLMs) enthusiasts for the first Fixie Hackathon. Fixie.ai is a new platform for building, hosting, and scaling Smart Agents that extend the power of LLMs with connections to personal data sources, systems, and tools. The hackathon participants were some of the first people to get access to the Fixie platform and for an entire day, they built their own AI-powered Agents and integrate them with different APIs or data streams.
Writing your first agent in Fixie can be as simple as writing just one instruction sentence in English such as:
“ I am an intelligent and playful agent that can play the 20 questions game. I (the agent) will think of a random thing, and you (the user) can ask me questions and try to guess what I’m thinking of ”.
These instruction sentences are also called prompts and they became the main form of interaction with LLM applications for generating text, images, music, or data charts. The beauty of the Fixie ecosystem is that it allows users to gradually make their agents more advanced by adding code prompts next to English text prompts. For example, you can use our built-in Python functions to make an agent that can search information from specific corpora, or write your own Python functions and combine them with the text prompts.
Another useful functionality is that your agents can talk to other agents and collaborate to solve specific tasks. For example, I can have a simple Haiku Art agent that generates both text and images by calling the Stable Diffusion Agent. The Fixie platform allows users to discover dynamically which agents need to collaborate to answer more complex questions. For example, if I ask what is on my calendar for tomorrow, the Calendar, the Time, and the Summarize agents come together to respond.
Our Hackathon participants really took advantage of the simplicity and the compositionality of the Fixie ecosystem and built over 70 agents during the event. I will highlight below the projects that were awarded by popular vote during the event.
The favorite demo award went to the Personal Email Assistant team. They built an agent that could sort your Gmail inbox and generate answers to new emails which you can review and edit before sending them (their project code is here).
The most creative demo award went to Nishchay who built a Movie Scene Agent which could find specific songs from movies based on a text description of a movie scene. He used the Spotify API for the song search.
The most ready-to-ship award went to the Youtube Recommender Agent, from Elaheh, Dhruv, Bahadir, and Adhyyan, who built an agent that could recommend movies based on specific topics.
Last, but not least the honorable mention went to David who built a Decomposer Agent to solve overly complicated tasks with lots of steps. His agent could let you answer fun queries such as: “ How many size 12 shoes would it take to fill up the whole volume of the Atlantic ocean?”
If these examples inspired you to go build your own Fixie Agent you can sign up for early access to our platform here and join our Discord server to get more ideas for new agents and stay in the loop for the latest Fixie updates.
