Agent Handoff SDK

Agent Handoff SDK addresses the challenge of transferring ongoing human work sessions to AI agents by providing a structured, machine-readable handoff mechanism. The tool introduces resume tokens, which encapsulate the human user's session context—including open browser tabs, scroll positions, video timestamps, decision threads, partial conclusions, and intended next steps—so that an AI agent can pick up the work exactly where the human left off, rather than starting from scratch. This approach is designed to prevent "cold-start" failures, where agents lack the context of prior human activity and reasoning.

The SDK is built around four core primitives: Session (capturing the work context on the device), Resume Token (a signed, scoped, time-bound credential granting agent access to the session), Resume Context (the structured payload delivered to the agent), and Reasoning Thread (a structured representation of partial conclusions, decision stance, blockers, and next-step intent). Resume tokens are implemented as short-lived, audience-bound JWTs, ensuring that access is limited to a specific agent or model. Redaction policies are enforced server-side so that agents receive only the content they are permitted to access, and every session hydration is auditable, recording who accessed what and when.

Agent Handoff SDK supports integration with multiple agent frameworks and ecosystems. It provides a consistent Resume Context schema that works across Anthropic Claude, Google Gemini, OpenAI Agents, and LangGraph. The SDK can be installed using npm or pip, and offers both REST and Model Context Protocol (MCP) API surfaces. Webhooks are available to subscribe to events such as session sharing and handoff. The Resume Context can be serialized into various formats suitable for different agent architectures, including Claude system prompts, Gemini system instructions, OpenAI Agents SDK inputs, and MCP resources.

The tool is intended for developers building agentic workflows that require seamless transfer of work between humans and AI agents. Its design emphasizes audience-bound access, structured reasoning representation, redaction, and auditability. Reference implementations and a full specification are available for those seeking to integrate the SDK into their own projects.