Skip to content
Hussh
Connect MCP

One Goal Framework — governed action execution

Public product and conceptual overview of the One Goal Framework, which powers intent execution behind Agent One and One Voice.

Concept

TL;DR: One Goal is Hussh's canonical Action-First framework that parses natural language intent across voice, chat, search, and command-bar inputs, maps them to cryptographically governed action contracts, and coordinates execution via approved service adapters.

Status as of 2026-07-06: see body.

Introduction to One Goal

The One Goal Framework is the canonical action-execution layer behind Hussh's Agent One. Rather than running as a generic chatbot or relying on unstructured model-driven tool usage, One Goal prioritizes structured intent mapping and absolute safety.

The core premise is Agent First:

  • Intelligence Layer: High-level system instructions, active application states, generated API contracts, and ADK/A2A-style proposals shape the interaction.
  • Contract Governance: The head agent (One) delegates to specific specialists (Kai, Nav, KYC) only through governed, cryptographically signed contracts.
  • Dynamic Planning: Intent is mapped to reusable resolvers and approved service adapters rather than hard-coding individual user stories.

Core Workflow & Components

  • Goal Planner (planOneGoal): Analyzes natural language intent and active UI/route states, ranks available action contracts, and resolves missing slot values. If information is missing, it prompts the user for exactly the next blocking input.
  • Goal Runner (runOneGoal): Instantiates a shared goal session, tracks progress across multiple steps, coordinates milestones, and handles cancellations.
  • Gateway Executor: Provides the absolute execution authority, bridging the goal runner to underlying system API endpoints and route settlements.

Privacy & User Control

The One Goal Framework enforces strict boundaries between public/provider interactions and private vault assets. Real-time models (such as Gemini Live) are used solely as proposal signals; they never execute tools or inspect plaintext credentials. Full execution authority remains local and governed, keeping private documents, decrypted knowledge models, and vault keys completely secure on-device.

Sources