Lets build a personalised agent harness in Safe Scala
Session Abstract
Inspired by Pi agent harness and TACIT. We will extend a core agentic harness. Provide agentic capabilities via the SKILLS.md standard with a twist: all tool calls are scala code execution of code generated by the agent itself. Sandboxing is provided by the new safe language mode. Come to learn and lets push it further.
Session Description
This session has two aspects, either someone wants to learn how agentic harnesses work, and this is a tutorial for how to help them and maybe send them on their own journey. But also I would invite a sort of hackathon where we really try to push what the agent can do.
This agent has only one tool it can use – execute scala code. the harness itself provides a thin communication channel that tool execution can use to communicate back to the harness. untrusted code is compiled under safe mode. Skills provided by the user come with Scala code that is dynamically loaded by the harness. The harness enforces Skill safety by compiling the provided interface under safe mode, and requires that any implementation must be safe to consume from safe mode.
Workshop (2 days)
