Ahoy - LLMs Enacting Interaction Protocols

Research Problem

LLM-based agents often require custom programming to participate in structured multiagent interactions. Ahoy investigates how agents can autonomously select and enact declarative interaction protocols while preserving protocol correctness.

Research Contribution

Ahoy introduces an LLM-based architecture for programming-free enactment of multiple BSPL interaction protocols. It separates language-model reasoning from protocol enforcement, allowing agents to adapt to different interaction structures without modifying their underlying source code.

Methodology and Evaluation

• The system combines LLM-based action selection with a BSPL protocol adapter that tracks state, validates actions, and handles external events.
• Across six evaluation tasks, Ahoy achieved 100% outcome success and correctness, compared with 33% outcome success for an LLM-only baseline.

Technologies

Large language models, BSPL, multiagent systems, protocol reasoning, protocol adapters, Python, FastAPI, Anthropic API.