Designing Atomicwork’s customized agentic framework for enterprise needs

This article is co-authored by Arkajit Datta and Riya Sebastian.

AI agents are reshaping workplaces, promising to transform how we work. But for enterprises, it’s not just about deploying AI—it’s about integrating agents that adapt, scale, and deliver in complex, ever-changing environments. While traditional workflows and frameworks address straightforward problems, they often fall short of meeting the dynamic needs of enterprises.

Enterprise service management requires more than clever automation. It calls for intelligence that evolves with the workplace—proactive, context-aware, and scalable. Agents must anticipate, decide, and act autonomously to manage advanced workflows and navigate sprawling SaaS ecosystems. This level of configurability and reliability is beyond the scope of most existing frameworks.

At Atomicwork, we saw this challenge as an opportunity. Rather than retrofitting generic solutions into enterprise molds, we built a custom framework fine-tuned for enterprise service management. Tailored for speed, and intelligence, it enables us to develop, deploy, and scale AI agents that meet enterprise needs without the compromises of traditional approaches.

In this article, we’ll take you behind the scenes of how Atomicwork’s AI agents came to life.

Setting the stage: Why off-the-shelf frameworks don't work

We began by evaluating existing frameworks like LangFlow, LangGraph, and SuperAGI—promising tools for AI agent and workflow development. However, these frameworks, while excellent for proofs of concept, lacked the reliability and customization needed for enterprise service management.

Our evaluation revealed several critical challenges like the frameworks being/having:

1. Not production-ready: Current off-the-shelf frameworks are still in their infancy, evolving rapidly but lacking the production-grade reliability enterprises demand. Their tight integration within proprietary ecosystems—such as LangChain’s integration with LangSmith for observability— also creates dependency risks, limiting long-term adaptability.
2. Too complex to customize: Frameworks like LangGraph require extensive custom coding and fragmented setups for advanced enterprise use cases. With a custom framework, we could consolidate these processes into a streamlined system, reducing maintenance overhead and improving visibility.
3. Limited configurability: Existing frameworks force businesses to adapt their needs to the tool, rather than the other way around. Our custom framework delivers unmatched configurability for ITSM and ESM workflows, allowing organizations to design scalable solutions aligned with their needs.

To address these challenges, we needed a solution purpose-built for scale and precision. Our custom in-house framework gave us complete control over every component, enabling us to tackle immediate issues while building a foundation for continuous innovation and long-term adaptability.

Our blueprint for multi-agent intelligence

Drawing inspiration from existing frameworks, our architecture is designed to enable the creation and orchestration of AI agents and workflows using modular components. Each workflow is structured as a graph, where vertices represent building blocks connected to form dynamic flows. Every vertex can have numerous components associated with it including agents, tools, function calls, or input/output operations. As a whole, this system will allow specialized agents to collaborate on a single problem and resolve it efficiently.

Agents: Agents are specialized modules responsible for reasoning, decision-making, or task execution. Each agent is equipped with specific instructions, predefined goals, and access to tools or APIs to accomplish its task. For example, a troubleshooting agent may decide to run diagnostics or reset a configuration based on user input.

Tools: Tools are operational components that execute specific actions, such as collecting system data, running commands, or executing processes on user devices. For instance, a Bluetooth troubleshooting tool might retrieve device status or turn Bluetooth on or off as required.

Function calls: Function calls allow workflows to integrate advanced logic such as calculations or classifications or access external services by invoking API integrations.

Input/output components: These components manage data flow in and out of the system, such as receiving user queries or delivering answers back to the user.

By connecting vertices to each other, the framework supports multi-step, multi-agent workflows capable of handling complex tasks. Each flow can involve multiple agents operating simultaneously, passing data or decisions back and forth to ensure the problem is addressed holistically.

For example, one agent might gather diagnostic data while another interprets it and a third agent takes action. This interplay of agents—each specialized, yet collaborative—ensures tasks are completed efficiently, accurately, and with minimal human intervention.

The modular design also ensures components are reusable, scalable and adaptable, allowing workflows to evolve as enterprise needs change. By designing each vertex to focus on a specific task or goal, the framework maintains clarity, avoids redundancy when scaling, and ensures optimal performance.

Improving user experiences with the agentic framework

Let’s break down how this framework would operate for an IT support use case.

When a user messages Atom saying “My Bluetooth headphones won’t connect”, the framework immediately kicks into action. The Classifier agent evaluates the query and assigns it to the Bluetooth agent, designed to handle such issues.

The Bluetooth agent invokes the on-device app to gather system metadata—checking if Bluetooth is enabled or if other devices are connected. It reasons dynamically, toggling Bluetooth, resetting the connection, or running diagnostics to identify hardware or software conflicts.

The Exit agent ensures a smooth handoff to Atom based on the defined threshold for user frustration, collating all gathered context and tracking user frustration levels to provide a final resolution or escalate further if necessary.

How Atomicwork’s framework is different

• Agent customization: Each agent is designed as a modular, goal-oriented, customizable component. Agents can reason and act independently or collaboratively, tailored to specific enterprise service management use cases.
• Tight integration with a custom on-device app: Our agents interface with a custom-built local application installed on user and enterprise devices. This tight coupling enables real-time metadata retrieval and command execution for effective troubleshooting and decision-making.
• Parallel execution: The framework supports parallel invocation of agents on the same query, to resolve multi-layered issues faster. For example, if a system issue involves both network and hardware components, the respective agents can work simultaneously, drastically reducing resolution time.
• Built-in security: Before any query reaches the framework, Atom’s guardrails, content moderation, and validation processes ensure only safe and valid queries are processed.
• Scalability and performance: By leveraging faster, task-specific LLMs, our framework minimizes response times while maximizing efficiency. It also consolidates outputs from multiple agents to make informed decisions.

Being future-ready for enterprise growth

Atomicwork’s multi-agent framework isn’t just solving today’s challenges—it’s a foundation built to meet the evolving demands of tomorrow’s enterprises. Designed for agility, it will soon enable us to deploy agents for different use cases across teams and industries faster. Imagine an IT manager effortlessly provisioning secure access across global teams or automating compliance checks for hybrid infrastructures—all without engineering overhead.

As the framework evolves, it will empower enterprises to navigate and support even the most demanding environments with ease. This isn’t just a tool—it’s a testament to Atomicwork’s commitment to innovation, transforming service management into a dynamic, intelligent process that allows enterprises to focus on what truly matters: driving growth and delivering value.

If you're interested to learn more on how we're thinking about unlocking AI for service management, drop a hi to our team here :)