Multi-Agent Framework (MAF)

The Multi-Agent Framework (MAF) in DAPPOS represents a groundbreaking evolution in AI orchestration for Web3, guaranteeing explainable AI decisioning and production-grade reliability.

Stateful Graph Backbone: The Foundation of Composable Autonomy

At the core of MAF lies the StateGraph, a structured state schema composed of a collection of variables that forms the backbone for all operations. The StateGraph encapsulates key system elements such as messages for inter-agent communication, plans for structured workflows, artifacts (e.g., data outputs or models), UI interrupts for human oversight, final outputs, errors for robust handling, and flags for conditional logic. These variables are passed seamlessly across each state in the system, serving as the decision-making foundation for controller nodes at each stage.

In MAF, each node represents an AI agent for a designed task. Specifically, controller nodes evaluate the current state to determine the next flow direction and update the state variables accordingly. The process of a user query begins with the main controller task_planner_node, which analyzes user intents to classify the task type—such as information gathering, strategy execution, or alpha detection—and deterministically routes it to the appropriate subgraph (search, DeFi, or opportunity).

Within each subgraph, specialized controller nodes handle internal flows, ensuring predictable paths that minimize variability and enhance trust in dynamic Web3 environments. This setup powers deterministic controllers and AI interrupt mechanisms, preventing errors while allowing freedom in AI decision-making.

Subgraphs: Specialized Modules for Web3 Intelligence

MAF's subgraphs are modular building blocks, each tailored to specific Web3 domains with deterministic node flows that process data efficiently. These subgraphs leverage the stateful backbone to integrate seamlessly, enabling adaptive intelligence across trading, analysis, and execution. For example:

• Search Subgraph: This subgraph combines autonomous controllers with enhanced review cycles for continuous deepening and refinement, embodying an agentic search loop. The flow starts with search_controller_node that selects and parallelizes agents (e.g., google_search_agent_node for web data, twitter_search_agent_node for social sentiment, parallel_search_coordinator_node for multi-source queries). Outputs are merged in merge_results_node, filtered for relevance, and refined through an iterative enhanced review loop until confidence criteria (e.g., source diversity or accuracy scores) are met, which is checked by enhanced_search_review_node. The summarize_node then produces cited answers with frontend-friendly markers, ensuring transparency and usability.

Search Subgraph

• DeFi Subgraph: Centered on tool-augmented execution, this subgraph generates and validates DeFi strategies using over 200 integrated tools for precise operations like API queries or trade simulations. It begins with fetch_market_data to retrieve real-time blockchain metrics, followed by strategy_generator to create optimized plans (e.g., yield farming or liquidity provision). The strategy_validator closes the loop by assessing risks, simulating outcomes, and incorporating optional user choices.

DeFi Subgraph

• Opportunity Subgraph: This subgraph identifies high-potential Web3 opportunities by scanning markets, trends, and signals. It flows through opportunity_search_node for initial detection, fetch_market_data for validation, and an opportunity_node that synthesizes insights into actionable recommendations, such as emerging meme coins or protocol upgrades.

Opportunity Subgraph

Advanced Features: Enhancing Collaboration, Execution, and Interactivity

Building on the core subgraphs, MAF incorporates advanced capabilities to foster human–AI synergy and real-time adaptability.

• Multi-Agent Search: Central to the search subgraph, this feature enables autonomous and collaborative intelligence, where a controller dynamically selects agents based on query complexity for parallel execution, boosting speed and breadth. For example, evaluating a meme coin's potential might invoke Google agents for news, Twitter agents for sentiment, and parallel custom agents for on-chain data. The enhanced review loop iterates refinements, and the summarizer outputs structured, cited responses optimized for frontend display, with HCI interrupts for real-time oversight in mission-critical actions.

• Tool-Augmented Plan Generation: Particularly prominent in DeFi and opportunity subgraphs, this integrates structured tools to transform abstract insights into executable strategies. Validator nodes ensure safety by cross-checking against risk parameters, with optional UI interrupts for user approvals, closing the intelligence-action loop and preparing plans for seamless handover to the Execution Layer.

• Streaming and Interrupts: To support interactive experiences, MAF streams messages during long-running tasks, keeping users informed in real-time. When user queries or provided information are insufficient for accurate analysis, the AI pauses execution to seek clarification by prompting the user with targeted questions, ensuring completeness and reliability.

In essence, DAPPOS's Multi-Agent Framework redefines Web3 intelligence by combining composable autonomy with safety rails, enabling users to tackle complex tasks with unprecedented precision and adaptability. This framework not only outperforms generic AIs in domain-specific reasoning but also paves the way for collaborative, evolving ecosystems where AI and humans co-create value.