Challenges I Ran Into

Building the agent presented several technical and architectural challenges, primarily around designing a fully agentic AI that could safely handle private ZEC transactions. The core hurdles included:

1. Orchestrating LangGraph with NEAR AI

   • Problem: NEAR AI provides a powerful inference backend, but it doesn’t natively handle complex, multi-step decision flows. I needed the AI to not just respond to prompts but to autonomously execute multi-stage actions—evaluating fundraisers, computing trust scores, and preparing cross-chain transactions.
   • Solution: I wrapped the NEAR AI model inside a LangGraph agent, which allowed me to define stateful, conditional execution paths. Each node in the graph represented a sub-task—like "fetch fundraiser metadata," "compute risk score," or "prepare TEE transaction payload." LangGraph handled the orchestration, ensuring outputs from one step fed correctly into the next, while also allowing asynchronous and parallel execution for efficiency.

2. Building the Trust Score Engine

   • Problem: Evaluating causes required combining on-chain data, off-chain metadata, and social trust signals, all while keeping the donor’s intentions private. There was a risk of leaking sensitive information if the AI queried external APIs carelessly.
   • Solution: I developed a 15-signal Trust Score Engine running inside TEE enclaves. Signals included project history, verification status, social proof, and historical donation patterns. The engine aggregates these into a single score, which the agent then uses to prioritize fundraisers. By keeping all computations inside the TEE, no user or transaction data ever leaves the shielded environment, solving both privacy and reliability concerns.

3. Complex Orchestration Flow

   • Problem: Coordinating the entire pipeline—from user intent → LangGraph agent → Trust Score evaluation → TEE execution → cross-chain settlement—was challenging due to dependencies between asynchronous steps.
   • Solution: I implemented a modular orchestration design within LangGraph, where each step’s inputs and outputs are explicitly defined, allowing easy debugging and monitoring. This modularity also enabled testing individual components in isolation before connecting the full flow, which reduced bugs significantly.

4. Testing and Debugging in a Privacy-First, Fund-Sensitive Environment

   • Problem: Verifying the full transaction flows required running real cross-chain ZEC transactions, which was costly and risked wasting funds during iterative testing. Standard debugging inside TEEs was limited, making it hard to ensure correctness without actually moving value on-chain.
   • Solution: I developed a safe simulation layer and cryptographically verifiable audit framework. This allowed me to replay full transaction flows and validate logic using minimal test funds, while producing anonymized, provable traces of every step. As a result, I could rigorously test and debug the system without exposing sensitive data or unnecessarily spending real ZEC.

Outcome: By combining LangGraph for orchestration, NEAR AI for agentic intelligence, and TEE-based execution for privacy, I successfully built a pipeline where donors can simply say:

"Donate 5 ZEC to high-trust privacy NGOs in Africa,"
and the agent autonomously evaluates options, computes trust, and executes a private, cross-chain transaction, all verifiable and secure.

This approach addressed both complex agent orchestration and privacy-preserving AI decision-making, demonstrating a first-of-its-kind TEE-powered agent for ZEC philanthropy.