Developing Agents: From Magic to Engineering

The Story

*Please RSVP ahead of time and bring a government-issued ID for check-in.* We use AI chats and coding tools every day. This session explores what happens when you call an LLM from your own code to make decisions and process data. Your program needs more than free-form text: it needs values it can branch on and functions it can run. We start with bare `curl` requests, moving from free text to structured output to tool calls. You’ll see what a tool call actually is: structured data naming a function and its arguments, with your code responsible for executing it. Then we hand that loop to [Pydantic AI](https://ai.pydantic.dev/), turn a plain Python function into a tool, and introduce a few basic multi-agent patterns. Finally, we introduce [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev), a new [System One](https://en.wikipedia.org/wiki/Thinking,_Fast_and_Slow#Two_systems) decision model that makes quick decisions in a type-safe way. We use it in two roles: deciding which tool to call, and grading the results as an eval judge. You’ll leave with a clear picture of how LLMs are integrated into a modern software stack, what it costs, and how to test whether it works. **About the Speaker** [Adam Matan](https://linkedin.com/in/adamatan) is the co-founder and CTO at [Breana.ai](http://breana.ai/) and previously led AI teams at Compass. **Venue Sponsor:** CIC Cambridge [https://cic.com/about/](https://cic.com/about/)

Get There

Similar Events