Session Description

In this talk, we will introduce a project shared between EPFL and VirtusLab focused on building a production-grade, permanent AI agent based on the TACIT tool developed by Martin Odersky’s team.

https://github.com/lampepfl/capybaraclaw

The talk will be composed of:

1) Introduction: We will explain what the landscape of permanent agents (like OpenClaw) looks like, what security and safety issues exist within these projects, and how that limits their implementation in serious contexts.
2) Analysis of the problem space: we will address these weaknesses by decomposing the issue to find what the nature of the problem is – generality and elasticity vs severe limitations imposed by sandboxing, what solutions are proposed by AI companies, and finally we’ll describe how TACIT and CapybaraClaw solve whole categories of these problems by using Scala’s new capture checking facilities to enforce rules while preserving utility.
3) Deep dive: We will walk through the agent’s architecture and capabilities in detail. The focus of this part is to show how Martin Odersky’s new approach unlocks stronger safety guarantees for AI agents.
4) Summary: We’ll talk about the roadmap of the project and our goals, share links and source materials (whitepapers, CTF repos where Frontier AI was used to find exploits against TACIT, etc.).

What the audience will learn: The primary benefit for attendees is the in-depth explanation of how the general and abstract system of Scala 3 capture checking and capabilities can be used productively in a use case that is much different from the usual coding.

The presenters for this talk may still change!