Rootconf 2026 Annual Conference

Session title

Beyond the AI Agent Demo: Engineering Reliable Enterprise Workflows with Graph-Based Orchestration

One-line summary

What happens when AI agents make unpredictable decisions in business-critical workflows? We share how we engineered a production multi-agent platform using graph-based orchestration, persistent state, event-driven architecture, and human-in-the-loop controls to make enterprise AI workflows more reliable.

What problem are you addressing?

Building an AI agent that can reason through a task is relatively straightforward. Ensuring it executes a complex enterprise workflow reliably, securely, and predictably is much harder.

Our Digital Workplace team built a multi-agent platform to automate laptop lifecycle workflows involving multiple enterprise systems, evolving business policies, asynchronous operations, and mandatory human approvals.

Our initial approach relied on agent-driven orchestration. During implementation, we encountered challenges with unpredictable execution paths and the possibility of mandatory workflow steps or human approvals being skipped. These behaviours are unacceptable in enterprise workflows where execution must be controlled, traceable, and repeatable.

We addressed these challenges by implementing graph-based workflows alongside persistent state management using Cloud Firestore, event-driven coordination, and integrations through the Model Context Protocol (MCP). The platform is live and has successfully completed 782 tasks.

This experience highlights an important engineering question: how do we combine the flexibility of probabilistic AI reasoning with the predictability and control of deterministic software systems?

Who is the intended audience?

Platform Engineering, SRE, DevOps, Infrastructure, Backend Developers, AI Engineers, Cloud Architects, and Engineering Leaders.

Level

Intermediate to Advanced

List one or two practical takeaways

1. How to decide which parts of an enterprise workflow should use AI reasoning and which should remain under deterministic workflow control.
2. How to design long-running agentic workflows using graph-based orchestration, persistent state, event-driven communication, and explicit human-in-the-loop steps.

What will you share?

• Architecture and design decisions: Our multi-agent architecture using Google ADK and Vertex AI, with an orchestration layer and specialized agents.
• Implementation details: How graph-based workflows and Cloud Firestore persistence support controlled execution and long-running processes.
• Production experience: Lessons from a live platform that has successfully completed 782 tasks.
• Failure modes and debugging: The limitations of relying on LLM reasoning to control mandatory workflow steps and approval gates.
• Enterprise integration: How MCP provides a standardized boundary between agents and enterprise systems.
• Operational challenges: Managing asynchronous human interactions, workflow state, retries, security requirements, and operational visibility.
• Trade-offs and alternatives: Agent-driven orchestration versus deterministic workflows, flexible reasoning versus predictable execution, and the role of humans in automated processes.

What is your experience with this problem?

• Production system
• Internal engineering project
• Hard-earned engineering lesson

What approaches failed, disappointed, or created unexpected problems?

Relying on LLM-driven reasoning to control the entire workflow proved insufficient for business-critical execution. Agents could produce inconsistent execution paths and potentially skip mandatory steps or human approvals.

We also encountered the challenge of managing workflows that span multiple interactions and depend on human actions. An agent’s in-memory context alone is insufficient for reliably managing processes that may pause and resume much later.

Security considerations also influenced our design. Capturing sensitive employee information directly through conversational interfaces raised concerns, requiring us to reconsider how sensitive information should be collected and handled through enterprise systems.

These experiences reinforced that prompt design alone cannot replace explicit workflow controls, durable state, and well-defined security boundaries.

What will you do differently today?

We would establish deterministic workflow boundaries earlier in the design process.

Rather than allowing an LLM to control every stage, we would use AI reasoning for contextual interpretation and decisions that benefit from flexibility, while enforcing mandatory steps, approvals, and state transitions through graph-based workflows.

We would also treat persistence, human-in-the-loop interactions, security boundaries, and operational visibility as core architectural requirements from the beginning rather than capabilities to add later.

What trade-offs did you consider?

• Autonomy versus control: Greater agent autonomy offers flexibility but makes execution less predictable. We chose explicit workflow control for mandatory business operations while retaining AI reasoning where it adds value.
• Agentic orchestration versus graph-based workflows: Agent-driven coordination can adapt to context, while graph-based orchestration provides more explicit execution paths and control over required steps.
• In-memory context versus persistent state: Externalizing state to Cloud Firestore adds storage and integration complexity but enables long-running workflows to pause and resume reliably.
• Direct integrations versus MCP: An MCP integration layer introduces an additional component but helps separate agent reasoning from enterprise system interactions.
• Automation versus human intervention: Fully automated execution may reduce manual effort, but explicit human approvals remain necessary for selected decisions, exceptions, and sensitive operations.

Our approach prioritizes predictable execution, security, and operational reliability without eliminating the benefits of AI-driven reasoning.

How can this help other practitioners?

This experience provides a practical design approach for teams moving agentic AI from prototypes into production.

Practitioners can use these lessons to identify where AI autonomy is appropriate, prevent LLM reasoning from becoming the sole controller of critical business processes, design durable workflows that span multiple interactions, and incorporate human approvals and security boundaries into the architecture.

The broader lesson is that reliable agentic systems require more than capable models. They need the same engineering discipline applied to distributed systems: explicit state, controlled transitions, well-defined integration boundaries, failure handling, and operational visibility.

Current state

Production experience

The platform is live and running successfully. Cloud Firestore persistence and graph-based workflows are implemented, with 782 successfully completed tasks.

Tags

#agents #platformengineering #reliability #devops #architecture #mcp #cloudnative #failurestory #casestudy #security #governance