Research / Agentic Systems

Building agents that can operate with structure.

Agentic software raises engineering questions that conventional chatbot interfaces do not resolve on their own.

Open questions

What we're asking

How should an agent receive permission to act?

How can execution be audited?

What happens when a task fails midway through a workflow?

How should memory, tools, and multiple runtimes interact?

These questions are explored in practice through the Karnelian Harness.