Harn
A programming language and runtime for building AI agents that manages model calls, tool execution, permissions and history, so the same agent code can run against local or hosted models.
Harn is a language and runtime purpose-built for defining an agent's task, prompts, tools and rules, while Harn itself runs the conversation between the model and the tools, enforces permissions on what each tool may access, and saves a record of the work. It normalizes differences between model providers so the same agent code can target either local or hosted models without change.
Harn is meant for agents that take several steps, such as investigating a failed job or reviewing a change, and ships bundled demo agents that replay recorded model responses so the workflow can be tried without any API key or provider credentials. It installs through a shell script installer that verifies release checksums, and the project describes itself as pre-1.0, with its language, standard library and CLI still subject to change.