AI Agent Framework

How Eliza and LangGraph orchestrate the Metalos assistant across portfolio, governance, and automation tasks.

Metalos uses a multi-node architecture powered by Eliza and Langgraph that combines specialized workflow nodes for different tasks. The Metalos agentic infra handles portfolio analysis, research, proposal generation, and data retrieval. The system combines deterministic workflow graphs, carefully engineered prompts, and contextual data to deliver intelligent responses.

High-Level Architecture

The AI layer sits between user-facing apps and the underlying automation services. When a user opens the portfolio chat window, the client sends the current wallet address, a snapshot of their portfolio, and the conversation history to a Metalos API endpoint. The server passes that bundle to a LangGraph StateGraph workflow. The workflow first verifies and enriches portfolio data, then invokes an LLM with a custom system prompt that includes vault descriptions, risk characteristics, and the user's actual balances. The assistant's response is streamed back to the client and displayed in the chat UI.

Phase 1 keeps data collection client-driven for simplicity. Phase 2 will move portfolio aggregation and risk analytics fully server-side so responses rely on Metalos-signed data rather than client-provided snapshots.

Two Agent System

Metalos uses two separate LangGraph agents for different use cases:

1. Portfolio Chat Agent (Basic)

Location: Embedded in Next.js app (lib/langgraph/portfolio-agent.ts)
Purpose: Portfolio-specific questions and analysis

Nodes:

fetch_portfolio - Loads wallet positions, normalizes balances, calculates weighted APY
chat - LLM with portfolio-aware system prompt

State: Conversation history, portfolio data, wallet address

Use Cases:

"What's my portfolio worth?"
"How is my Extremely Safe vault performing?"
"Should I rebalance?"

2. Research Agent (Advanced)

Location: ROFL agent service (rofl-agent/src/research-agent/)
Purpose: DeFi research, proposal generation, documentation retrieval

Nodes:

chat - Main conversational node with tool calling
proposal_generation - Guided governance proposal creation (multi-step flow)

Tools Integrated:

Data Sources: DexScreener, DeFiLlama, BaseScan, GitHub (22+ tools)
Deep Agent: File operations (read, write, edit, ls), todo management
RAG: Documentation retrieval from Supabase vector store
Proposal Utils: Template selection, validation, formatting

State Management:

Checkpointer: Supabase-based persistence for session recovery
Files: Virtual file system for research artifacts
Todos: Task tracking across sessions
Conversation History: Full context maintained

Use Cases:

"Analyze the AERO token on DexScreener"
"What's the TVL for Morpho on Base?"
"Create a proposal to add Aave V3 vault"
"Search the documentation for withdrawal information"

This is a two-agent architecture: simple portfolio agent for basic questions, comprehensive research agent for advanced analysis and proposal workflows.

Loading diagram...

Prompt Engineering

The system prompt generated in system-prompts.ts gives the LLM a complete snapshot of the user’s holdings:

Total portfolio value and weighted APY
Each vault position with live balance and APY
Canonical descriptions of available strategies (Extremely Safe, Not Very Risky, Moderately Risky)
Behavioral guidelines such as “explain impermanent loss if LP vaults are mentioned”

Because the prompt is regenerated every turn, the assistant always references the latest balances and strategy settings without needing persistent memory.

TEE-Protected Execution with Oasis ROFL

The Research Agent runs inside Oasis ROFL (Runtime Offchain Logic), a Trusted Execution Environment (TEE) that provides hardware-level security for sensitive AI workloads.

Why TEE for AI?

Code Integrity:

Cryptographic attestations prove which exact code executed
Measurement hash verifies official Metalos agent (not backdoored)
Community can audit which agent version produced results
Tamper detection via cryptographic signatures

Runtime Privacy:

Code and data encrypted in TEE CPU memory during execution
Cloud providers can't access running TEE processes (even with root)
Proprietary AI prompts can't be extracted from TEE memory
Secure processing environment

IP Protection (Metalos):

Proprietary prompts protected in TEE memory
Strategy algorithms can't be extracted
Model weights inaccessible to outsiders
Competitive advantage maintained

Unique Advantage: Metalos is one of the only DeFi platforms offering TEE-protected AI research. This enables confidential analysis that traditional AI services can't provide.

How It Works

Configuration

Metalos can be configured to run the Research Agent in TEE or standard mode via environment variables:

USE_TEE_AGENT=true → All queries use Oasis ROFL TEE
Not set → Standard execution (faster, no TEE protection)

Query Processing

When TEE is enabled:

Query sent via HTTPS to ROFL endpoint
Processed inside Oasis secure enclave
Code and data encrypted in TEE CPU memory during execution
Agent has access to all tools and features

Attestation Generated

TEE produces cryptographic proof:

Signature over response
Measurement hash of code
Timestamp and TEE identifier

Response with Proof

You receive:

AI response (decrypted)
TEE attestation (verifiable proof)
Shield icon (🛡️) indicating TEE execution

TEE usage is currently an all-or-nothing configuration, not per-query routing. When enabled, all Research Agent queries run in the TEE. When disabled, all queries use standard execution.

Viewing TEE Attestations

In the UI: Look for the shield icon (🛡️) next to AI responses

Click the shield to see:

Hardware TEE Verified status
Cryptographic signature
Timestamp and TEE identifier
Measurement hash (code verification)
App ID and version

See TEE Attestations Guide for detailed verification instructions.

TEE vs Standard Execution

Feature	TEE (ROFL)	Standard
Privacy	Maximum (hardware encrypted)	Standard HTTPS
Verifiability	Cryptographic attestations	None
Latency	+200-500ms (encryption overhead)	Fast
Cost	Higher per query	Lower
Configuration	`USE_TEE_AGENT=true`	Default

TEE is configured at deployment level (all queries or none), not per-query. When TEE is enabled, you get maximum privacy for all research with slightly higher latency.

Eliza Integration

The Eliza UI hosts the chat interface inside the portfolio page. It handles:

Rendering streaming responses and conversation history
Displaying portfolio cards alongside the chat so users can cross-check numbers
Triggering intent suggestions (e.g., “Deposit more into Extremely Safe”) that deep-link into the app
Passing the correct wallet address and portfolio snapshot to the agent API

Because Eliza is highly customizable, future enhancements can add voice input, multimodal responses, or embedded action buttons driven by agent output.

Implemented Features

Research Agent Capabilities (Live)

Data Source Integration:

✅ DexScreener: Token prices, pairs, liquidity, OHLCV data
✅ DeFiLlama: Protocol TVL, yields, historical performance
✅ BaseScan: On-chain transactions, contract verification, wallet analysis
✅ GitHub: Repository analysis, audit detection, commit activity

RAG Documentation Retrieval:

✅ Vector store with Metalos documentation embeddings
✅ Semantic search for platform knowledge
✅ Supabase pgvector integration
✅ Automatic context injection for platform questions

Proposal Generation:

✅ Multi-step guided workflow
✅ 5 proposal templates (Signaling, New Vault, Recategorization, Removal, Custom)
✅ Data validation with research tools
✅ Formatted output ready for governance

Deep Agent Tools:

✅ File system operations (read, write, edit, list)
✅ Todo management for research planning
✅ Session artifacts stored as structured files
✅ "Continue Research" capability via file sharing

State Persistence:

✅ Supabase checkpointer for session recovery
✅ Full conversation history maintained
✅ Files and todos persist across sessions
✅ Resumable research workflows

Portfolio Agent Capabilities (Live)

Portfolio Analysis:

✅ Real-time position data
✅ Vault-aware system prompts
✅ APY and balance calculations
✅ Risk-appropriate explanations

Conversational Support:

✅ Natural language queries
✅ Streaming responses
✅ Context-aware answers
✅ DeFi concept explanations

Roadmap: Future Enhancements

Planned Additions

Portfolio Agent Enhancements:

Autonomous portfolio analysis: Proactive optimization suggestions
Risk analytics workflows: Monte Carlo simulations and stress testing
Action routing: Prepare transaction batches from natural language

Research Agent Enhancements:

Additional data sources: Nansen, social sentiment, Dune Analytics
Advanced RAG: Search community research, not just docs
Cross-protocol analysis: Compare opportunities across DeFi ecosystem

System-Wide:

Automated rebalancing: Periodic portfolio rebalancing based on conditions
Governance copilot: Proposal impact analysis and voting recommendations
Multi-agent coordination: Portfolio and research agents share insights

Current Status: Both Portfolio Chat (basic 2-node) and Research Agent (advanced multi-node with RAG, proposal generation, and Deep Agent tools) are live and operational.