Frame Handshake Protocol (FHP) v1.0: A Fail-Closed Coordination Protocol for Agentic and Multi-Agent AI Systems
Kon Lionis · Zenodo (CERN European Organization for Nuclear Research) · 2026
This work introduces the Frame Handshake Protocol (FHP), a deterministic coordination layer for agentic, multi-agent, and human-routed AI systems. FHP addresses known failure modes in large language model ensembles — including stochastic drift, implicit frame shifts, role confusion, partial state loss, and state divergence — by formalising interaction as a fail-closed, protocolised process. The framework combines symbolic governance mechanisms, such as explicit frames, constraints, permissions, roles, recursion depth, and router authority, with distributed-systems-inspired primitives such as shared state tracking, lock integrity, relay resume cards, adaptation commits, recovery procedures, and halt conditions. Its purpose is to make model coordination more auditable, replayable, recoverable, and governable at the conversational layer. This v1.2 release includes the FHP Unified Operator Document, a single-file runtime document designed to be uploaded directly into an LLM. Rather than functioning only as a descriptive specification, the operator document is intended to be enacted: a model reading it should be able to begin a valid FHP session by announcing itself, proposing or negotiating a frame, entering an explicit lock, and continuing the task under agreed constraints. The document includes router instructions, packet templates, MYTHIC/SRP integration, adaptation logic, relay recovery rules, router alerts, and fail-closed halt behaviour. FHP is designed primarily for human-in-the-loop coordination, where the human acts as router or steward: confirming frames, carrying relay context between models, approving adaptation requests, monitoring drift or mismatch signals, and halting sessions when ambiguity or instability appears. This makes it suitable for structured human–LLM sessions, LLM-to-LLM relay experiments, symbolic alignment workflows, multi-model comparison, and controlled exploratory reasoning. FHP does not claim to be a truth engine, content filter, reasoning method, or replacement for safety policy. It is an interaction-layer governance protocol: it controls how a session is framed, locked, adapted, resumed, and halted, while preserving explicit human authority over continuation. An interactive browser-based visualiser is also provided as a public-facing companion artifact for the Frame Handshake Protocol, demonstrating session replay, shared state coordination, lock integrity, and routed control flow.