By James Aspinwall, co-written by Alfred Pennyworth (my trusted AI) — March 7, 2026, 12:56
The Missing Piece in Enterprise AI
Every enterprise building AI agents today hits the same wall. The models are capable. The tools exist. The APIs are documented. But the agents still hallucinate, lose context mid-workflow, and can’t explain why they reached a conclusion.
The problem isn’t intelligence. It’s memory.
AI agents need more than a flat pile of document chunks retrieved by semantic similarity. They need to understand how things relate — which supplier feeds which product line, which regulation governs which process, who owns which account and why it matters. They need a structured, queryable understanding of the business.
Arango builds exactly that layer. WorkingAgents orchestrates the agents that consume it. Together, they form a complete agentic AI stack: context plus action.
What Arango Actually Does
ArangoDB is a native multi-model database — not a wrapper around separate engines, but a single core that handles five data models in one system:
- Graph: Property graphs with vertices and edges, supporting multi-hop traversals and pattern matching
- Document: Flexible JSON documents with no pre-defined schema
- Key-Value: High-speed lookups for simple access patterns
- Vector Search: Semantic similarity queries on embeddings via FAISS integration
- Full-Text Search + Geospatial: Boolean retrieval with ranking, plus geographic queries
The critical point: all five models share a single query language (AQL) and run in the same engine. A single query can traverse a graph, filter documents, run vector similarity, and do full-text search. No data pipelines stitching separate databases together. No ETL jobs. No eventual consistency between systems.
Arango has rebranded to arango.ai and positions itself as the “Contextual Data Layer” for agentic AI — the bridge between raw enterprise data and AI applications that need structured business context. They’re a member of NVIDIA Inception, have 13,000+ GitHub stars, and serve Fortune 500 companies across financial services, healthcare, telecom, and government.
What WorkingAgents Does
WorkingAgents is an AI orchestration platform built on the Model Context Protocol (MCP). It connects AI models to tools, APIs, and workflows, turning natural language into coordinated multi-step actions. It’s model-agnostic — it works with Claude, GPT, Gemini, open-source models — and its consulting arm helps medium-size companies integrate AI into their existing operations.
WorkingAgents handles the doing. It coordinates agents that pull data from a CRM, draft communications, update records, trigger notifications, and chain these steps into coherent workflows. What it needs is a data backend that can give those agents deep, relationship-aware business context — not just document retrieval, but structured understanding.
Where the Synergy Lives
1. MCP Is the Shared Protocol
This isn’t a theoretical integration. Arango already ships a production MCP server — mcp-arangodb-async — with 46 tools across 11 categories: AQL query execution, collection management, graph operations, document sampling, analytics, backup/restore, and content conversion. It’s async-first Python, available on PyPI, and compatible with any MCP client.
WorkingAgents orchestrates MCP tools. Arango exposes its entire database as MCP tools. The connection is architecturally native. A WorkingAgents agent can query Arango’s knowledge graphs, traverse entity relationships, run hybrid vector+graph searches, and manage collections — all through standard MCP tool calls, with no custom integration code.
This is the cleanest possible integration path: both platforms already speak the same protocol.
2. Graph + Vector Solves the Hallucination Problem
Standard RAG with vector databases (Pinecone, Weaviate, Qdrant) retrieves document chunks by semantic similarity. It works for simple Q&A. But it loses structural relationships — who reports to whom, which product uses which component, which compliance rule governs which workflow. When an agent needs to reason across connected entities, vector-only RAG falls apart and the model fills gaps with hallucination.
Arango’s HybridGraphRAG combines three retrieval mechanisms in a single query:
- Vector search finds semantically similar content
- Graph traversal follows named relationships between entities
- Full-text search matches precise terminology
Published benchmarks show HybridGraphRAG achieves 8% higher factual correctness and 7% better context relevance versus standard vector RAG. Responses are more consistently grounded in retrieved context because the graph structure constrains what the model can plausibly say — relationships are explicit, not inferred.
For WorkingAgents, this means agents that make fewer mistakes, explain their reasoning through graph paths, and handle multi-step queries that span organizational boundaries.
3. Context-Aware Workflow Orchestration
Consider a WorkingAgents agent handling: “Prepare a risk assessment for the Meridian deal.”
With a vector database, the agent retrieves document chunks that mention “Meridian” and “risk.” It gets fragments. It might miss that Meridian’s primary supplier is under regulatory review, or that a similar deal in the same vertical churned last quarter.
With Arango as the data layer, the agent traverses a knowledge graph:
- Meridian → supplies_to → Product Line X → regulated_by → Compliance Framework Y (currently under review)
- Meridian → same_vertical_as → Oakridge → deal_outcome → churned Q3 2025 → reason → integration delays
- Meridian → key_contact → Sarah Chen → last_interaction → 45 days ago → sentiment → positive but cautious
The agent doesn’t just retrieve similar documents. It assembles a structured understanding of the deal’s context, risks, and relationship history. Then WorkingAgents orchestrates the output — drafting the risk assessment, attaching supporting evidence with graph-path citations, and routing it to the deal owner.
4. Knowledge Graph Construction from Existing Data
Most enterprises don’t have a knowledge graph. They have data scattered across CRMs, ERPs, shared drives, Slack, and email. Arango’s AI Suite includes automated knowledge graph construction — extracting entities and relationships from raw text using LLMs, grouping semantically similar entities, and storing the resulting graph.
WorkingAgents agents could trigger and manage this process: ingesting data from connected tools (Salesforce records, support tickets, contract documents), routing it through Arango’s graph construction pipeline, and keeping the knowledge graph current as business data changes. The orchestration layer feeds the data layer, which feeds the orchestration layer. A virtuous cycle.
5. Explainable AI for Regulated Industries
In financial services, healthcare, and government — industries where both Arango and WorkingAgents see demand — AI decisions need audit trails. “The model said so” doesn’t satisfy compliance.
Arango’s graph paths provide exactly this. When a WorkingAgents agent makes a recommendation, the reasoning chain is traceable: entity A is connected to entity B via relationship C, which triggered rule D, resulting in recommendation E. Every step is a named edge in the graph, not an opaque embedding similarity score.
This turns AI agent output from “probably correct” to “verifiably grounded” — a prerequisite for enterprise adoption in regulated sectors.
What a Joint Solution Looks Like
| Layer | Component | Role |
|---|---|---|
| User Interface | Natural language request | “Who are our at-risk accounts this quarter?” |
| Orchestration | WorkingAgents (MCP) | Breaks the request into steps, coordinates tool calls |
| Data Context | Arango (MCP server) | Graph traversal + vector search + full-text across enterprise knowledge graph |
| Action | WorkingAgents (MCP tools) | Drafts reports, updates CRM, notifies account managers |
| Audit | Arango graph paths | Explainable reasoning chain for compliance |
The Competitive Advantage of Combining
Neither platform alone covers the full stack:
- Arango without orchestration is a powerful database that still requires custom application code to connect AI models to business actions
- WorkingAgents without a contextual data layer orchestrates agents that retrieve flat document chunks and miss structural relationships
Together, they offer something no single vendor provides: AI agents that understand business context deeply (Arango’s graph + vector + document model), act on that understanding across enterprise tools (WorkingAgents’ MCP orchestration), and explain their reasoning through traceable graph paths.
The integration cost is near zero — both platforms already speak MCP. The value proposition is immediate — reduced hallucination, explainable decisions, and workflows that understand organizational relationships rather than treating the business as a bag of document fragments.
Next Steps for Partnership Exploration
- Technical proof-of-concept: Connect WorkingAgents to Arango’s MCP server, demonstrate a multi-step workflow that uses HybridGraphRAG for context retrieval and WorkingAgents for action execution
- Joint use case: Pick a vertical (financial services compliance or healthcare operations) and build a reference architecture showing the combined stack
- Co-marketing: “Context + Action” narrative for enterprise AI — Arango provides the brain, WorkingAgents provides the hands
- Shared MCP ecosystem: Contribute to the MCP tool ecosystem together, demonstrating interoperability as a selling point for enterprises evaluating agentic AI platforms
The thesis is simple: enterprise AI agents need structured knowledge and coordinated action. Arango owns the knowledge layer. WorkingAgents owns the action layer. MCP connects them cleanly. The companies that adopt both will build AI systems that actually work in production — not because the models are smarter, but because the data architecture gives them the right context to act on.