Multi-agent platform for enterprise workforce operations

Confidentiality: this page describes architecture patterns and functional areas only.

The problem

A large R&D organisation ran leave, transport, facilities, admin and HR requests through separate systems and manual approvals. Employees did not know where to go, managers approved by email, and operations teams handled exceptions by hand.

The goal was a single conversational entry point that could actually complete tasks across those systems — not a chatbot that pointed people at forms.

Constraints

• Actions had real consequences — leave balances, cab allocation, access management — so every agent needed clear authority limits and human approval at the right points.
• Corporate IT, security and data-privacy policies applied to every integration.
• Many concurrent users, and long-running conversations that had to keep context across sessions.

What I designed and built

1. Graph-based agent orchestration. A LangGraph supervisor routes requests to domain agents — leave, transport, admin, HR — each with its own tools exposed through MCP. AutoGen and Claude subagents handle sub-tasks that benefit from parallel execution.
2. Stateful agents with human-in-the-loop. Policy validation runs automatically; approvals pause the graph and resume on a manager's decision. State is persisted, so a request survives sessions and restarts.
3. Memory design. Short-term conversation memory per session and long-term memory per employee, so agents remember preferences and history without re-asking.
4. Model routing. Different models for different steps, chosen for cost, latency and quality — with evaluation harnesses to catch regressions when a model or prompt changed.
5. Delivery. Sprint planning, stakeholder management, risk and compliance sign-off, and the mobile and web front ends employees actually used.

Stack

LangGraph · AutoGen · Claude Agent SDK with subagents · MCP · Python · Node.js · React · Flutter · PostgreSQL · microservices
Models: Claude (Opus, Sonnet, Haiku) · OpenAI GPT · Llama · DeepSeek · Mistral

Outcomes

• Leave, transport, facility booking, inventory alerts, onboarding and HR queries all handled through one agentic interface, with automated policy checks and escalation.
• High request volumes handled through concurrent agent execution.
• A reusable pattern — supervisor graph, MCP tools, human-in-the-loop checkpoints, memory layers — that I now apply to every agent build.