@hashgraph namespace, requiring explicit import of needed modules.
Hedera launched version 4 of its Agent Kit on May 18, restructuring the developer toolkit into modular packages and introducing a new policy engine with four distinct lifecycle hooks to give developers more granular control over agent behavior.
"AI agents that interact with blockchain networks need guardrails," the official announcement stated, highlighting the need for developer-defined rules when agents can freely spend HBAR or mint tokens.
The new version breaks the previously monolithic hedera-agent-kit into smaller, scoped packages under the @hashgraph namespace. Developers now explicitly import plugins for core services like accounts, tokens, and consensus. The new policy engine allows developers to inject logic at four stages: Pre-Tool Execution, Post-Parameter Normalization, Post-Core Action, and Post-Tool Execution.
This shift forces developers to be more intentional about agent capabilities, potentially reducing bloat and improving security by ensuring an agent only has the permissions it strictly needs. The changes set the stage for more complex, production-ready AI agents on the Hedera network, with a new developer bounty program running from May 18 to June 21 to incentivize adoption.
The two primary themes of the v4 release are modularity and control. Instead of a single large package, the kit is now a family of @hashgraph-scoped packages. The core package contains shared types and the plugin system, while framework integrations for LangChain, Vercel AI SDK, and Google's Agent Development Kit (ADK) are in separate packages. This allows developers to install only the dependencies required for their specific use case.
A significant change is the move to explicit plugin imports. In v3, default tools could be loaded silently. Now, an empty plugins array means the agent has no tools, forcing developers to consciously opt into every capability, such as token minting or account queries.
The introduction of a hooks and policies system provides developers with fine-grained control over an agent's actions. The four lifecycle hooks enable logic to be inserted before a tool executes, after parameters are normalized, after the core action is formed but before it's sent on-chain, and after the tool has fully executed. This allows for the implementation of guardrails like spend limits, address allowlists, and detailed audit logging.
To support this, the way tools are built has been refactored. The new BaseTool abstract class replaces the previous plain object pattern, a tool's function into lifecycle stages that the policy engine can automatically hook into. While the old pattern is still supported for backward compatibility, new tools must use BaseTool to be governed by the policy system.
The update also includes support for Google's Agent Development Kit (ADK), enabling developers to build Hedera-powered agents using Google's framework and Gemini models.
This article is for informational purposes only and does not constitute investment advice.
