Content
High Dimensional Research announced that Nolita, their autonomous navigation stack, is undergoing reconstruction with a fresh Python codebase centered on "preliminary navigation planning and hierarchical agent delegation."
Northern Little Italy
The redesign emerged during Memory Index development. The organization prioritizes delivering extensible experiences for developers while avoiding proprietary black-box solutions. They advocate for treating models as "relatively agnostic slots within a stack, where you can use weaker or stronger models for the task as necessary."
Previous Architecture
The original Nolita system employed a monolithic agent approach:
- Users provide an objective and starting URL
- The system loops between itself and an LLM, progressing through pages
- The Memory Index substitutes similar historical actions when applicable
New Architecture: Navigation Hivemind
Rather than one agent, the redesigned system features three high-level agents and over a dozen low-level agents managed by an orchestrator:
- Planning Agent: Maintains strategic direction
- Execution Agent: Manages browser functions through specialized sub-agents
- Evaluation Agent: Assesses progress
"We no longer have to pass a full DOM tree analysis into context; we can preprocess it" with dedicated processing, potentially leveraging the Index for element analysis.
Why Python?
The team discovered superior orchestration capabilities in Python and found developers predominantly integrate Nolita within Python environments. The legacy TypeScript version will remain available with reduced functionality.
Preview
The new version is currently in private testing with design partners. Interested developers should contact Tynan at tynan.daly@hdr.is.