Capstone: NEXUS Deep Researcher

Canonical Source of Trust: Google Doc Tab Practice: The "Deep Agent" Capstone
Training Program: [Huy Chau] Generative AI Training Plan

---

NEXUS

**Timeline: **3.5 Weeks **Tech-stack: Python/Typescript, FastAPI/Fastify, LangGraph, LangFuse, LangChain, **Gemini models, Pydantic/Zod, llm-guard.

THE CORE DEEP RESEARCHER

**Timeline: **2 Weeks **Prerequisites: **Completion of Fundamental Plan Objective: Build the primary "intelligence collector" node. This agent must move beyond "chatting" to producing verified technical artifacts through systematic environmental scaffolding Specification:

• Technical Architecture & Harness
  • Orchestration Engine: The system must be built as a Directed Cyclic Graph (DCG) using LangGraph to support iterative feedback loops for error recovery
  • Agent Harness: Implement a "Deep Agent" harness that includes a Virtual File System (VFS). The agent must persist all intermediate research data to VFS files (e.g., research_notes.txt) rather than relying on the conversation history to prevent token overflow.
  • Compute Strategy: Deploy a "Reasoning Sandwich" baseline: utilize xhigh reasoning models (e.g., GPT, Gemini, or Claude Sonnet) for initial planning and final verification, while using standard models for implementation tasks
• Functional Requirements
  • Strategic Planning: Upon receipt of a query, the agent MUST first call a write_todos tool to decompose the objective into a structured checklist before executing any retrieval.
  • Tool Connectivity: Integration with external data (e.g., Tavily or local databases) must be executed strictly via the Model Context Protocol (MCP) using standardized clients.
  • Reliability Primitive: Implement a mandatory Build-Verify-Fix (BVF) loop. The agent must run a programmatic verification pass (e.g., a linter for code or a source-verification script for text) and is prohibited from terminating until all tests pass.
  • Durable Persistence: The system must utilize LangGraph Checkpointers to enable session resumption and failure recovery across long-running tasks
• Observability & Acceptance Criteria
  • Trace Fidelity: Every execution must generate a high-fidelity LangFuse trace demonstrating the full hierarchy of Runs, Traces, and Threads.
  • Performance Target: The agent must achieve a >90% Answer Faithfulness score as measured by automated RAGAS evaluators.
  • Safety Gate: The system must include a Human-in-the-Loop (HITL) interrupt that pauses execution for approval before the agent can finalize any permanent file writes

THE ENTERPRISE FEDERATED SWARM (EVOLUTION)

**Timeline: **1.5 Weeks **Prerequisite: **Completion of Beyond Plan & Observability Lifecycle **Objective: **Scale the Level 1 "Collector" into a secure, governed, and distributed ecosystem. You are now building the Supervisor and Security Layers that manage multiple specialized collectors Specification:

• Core Architectural Requirements
  • Heterogeneous Delegation: Upgrade the collector node to support the Agent-to-Agent (A2A) protocol. Implement a Supervisor pattern where a high-level manager (possibly built in Google ADK) delegates specialized sub-tasks to your LangGraph-based collector via standardized Agent Cards.
  • Agent Semantic Layer (L9): Implement protocol-level Semantic Handshakes (SL-HELLO, SL-SELECT) to negotiate and lock shared context definitions between the Supervisor and Workers before execution begins.
  • Shared Memory Layer: Transition from local VFS to a Shared Context Store (SCS) (Blackboard pattern), where all agents in the swarm read/write shared global states to reduce redundant central LLM orchestration calls
• Governance & Security
  • Managed Defense Layer: Deploy AWS Bedrock Guardrails or Meta LlamaFirewall at the ecosystem boundaries to enforce PII masking, toxicity filters, and Prompt Injection defense.
  • Behavioral Auditing: Integrate an AlignmentCheck mechanism that reasons over the entire ecosystem execution trace to halt the swarm if any worker's actions diverge from the primary user objective.
  • Mechanical Least Privilege: Ensure every agent identity in the swarm utilizes short-lived tokens (15–60 min) with automatic refresh to limit the blast radius of a potential compromise
  • Partial failure semantics and distributed consensus patterns
• Meta-Optimization (The Flywheel)
  • The Data Flywheel: The trainee must identify a reasoning failure in the multi-agent flow, capture the specific trace in LangFuse, and convert it into a permanent Regression Test.
  • Explainable HITL (XAI): Any high-impact action (e.g., final report publication) must pause for a Human-in-the-Loop approval gate, providing a "Feature Attribution" summary explaining why the system is confident in its decision
• Deliverables
  • Ecosystem Map: A complete architectural diagram showing A2A handshakes and SCS interactions.
  • Compliance Audit Log: A tamper-evident log documenting all tool calls, security blocks, and human validation steps.
  • Benchmark Report: A LangFuse report proving that the hardened Level 2 system maintains accuracy even under simulated adversarial attacks