Agentic Systems - Unlocking the Promise of Autonomous Execution of Complex Workflows

Credits for Input and Support: Veselin Pizurica and Leonard Donnelly of Waylay

Agentic systems show great promise in transforming how businesses manage complex workflows by breaking high-level goals into functional components, executing them autonomously, and adapting them in near real-time. While these systems are still evolving, their potential to revolutionize business operations is clear.

Step 1: From Goals to Functional Decomposition

The process begins with breaking down a high-level goal or prompt into a functional view. This step translates abstract business objectives—such as "optimize customer service operations"—into defined functions like issue categorization, response prioritization, and resolution management.

These functional components serve as the foundation for further decomposition into specific, executable tasks that can be mapped into a services architecture.

Step 2: Building the DAG from Functions, APIs, LLMs, and UIs

To achieve autonomous execution, agentic systems construct a Directed Acyclic Graph (DAG) that combines four core component categories: functions, APIs, LLMs, and UIs. Each of these plays a crucial role in executing tasks, and I’ve selected specific Greek letters to represent their unique roles:

·??? Λ (Lambda): Represents functions (not the same as functional), that can be either typed (strict enforcement of input and output expectations) or untyped. Lambda is traditionally associated with functions in programming, making it an appropriate representation for the functional units driving execution.

·??? Σ (Sigma): Represents APIs, which enable communication between services or inputs. APIs can also be typed or untyped, depending on the level of structure required for data exchange. Sigma symbolizes summation, reflecting the way APIs aggregate and transmit data across different services.

·??? Θ (Theta): Represents UIs (User Interfaces). Unlike functions or APIs, UIs must be flexible, adapting to user inputs and interactions. Theta, often representing complexity or unknown variables, fits well with the adaptable nature of user interface interactions that the system must autonomously handle.

·??? Δ (Delta): Represents Agents or Agentic Systems themselves, that provide delegated intelligence. Delta signifies change and adaptability, making it the perfect fit for LLMs that continuously learn and adjust to new inputs or conditions.

The DAG integrates these components, balancing the structured nature of functions and APIs with the flexibility of Agentic Systems and UIs. This structure ensures the system can adapt in near real-time while executing tasks efficiently.

Step 3: Custom LMs as the Foundation of a Language Model Operating System (LM OS)

At the core of agentic systems are custom Language Models (LMs), which provide the intelligence necessary to decompose, execute, and manage complex workflows. These LMs form the underpinning of what can be referred to as a Language Model Operating System (LM OS).

An LM OS integrates custom LMs customized to specific sectors, functions, and firms. These models allow agentic systems to not only break down goals into executable services available in an enterprise but also enable workflows that can continuously adapt, learn, and improve over time. The components of the LM OS include:

·??? Base LMs pre-trained with domain-specific knowledge, giving the system the context required to manage complex business functions.

·??? Large Action Models (LAM) that understand UIs, to enable autonomous execution of tasks that would typically require human intervention.

·??? Function LMs that understand the context and syntax for the use of functions as described in the Enterprise Function Library.

·??? APIs LMs that understand the context and syntax of APIs as cataloged in the enterprise API Library. ?

The LM OS can leverage the Base LM, integrated with the Λ, Σ, Θ, and Δ systems described above, and acts as the operational layer that dynamically coordinates workflows, learning from interactions, and adapting and evolving its capabilities, providing a robust framework for managing complex goals in a business environment.

Step 4: Autonomous Execution, Adaptation, and Remediation

Once the DAG is built and powered by components of the LM OS, it can move into the execution phase. During this phase, the system performs more than just task execution—it actively monitors and adapts to evolving conditions.

Autocorrection and Real-Time Remediation with ITSM

As the system executes the DAG, it continuously monitors for errors. If an issue arises, such as a data input mismatch or system error, the agentic system autonomously detects the error, queries the relevant IT Service Management (ITSM) system, and retrieves potential fixes. For example, it might look up error codes, bug reports, or documented solutions to autocorrect the issue without manual intervention.

Adapting to Evolving SOPs

In addition to error remediation, agentic systems dynamically adapt to changes in Standard Operating Procedures (SOPs). As new SOPs are introduced, the system updates its workflow, ensuring execution remains compliant with the latest standards. This adaptive capability allows the system to maintain efficiency and accuracy even as the rules evolve.

Multi-Agent Reinforcement Learning (MARL) and Multi-Agent Orchestration (MAO)

Agentic systems also excel in managing multiple agents working together to achieve complex goals. This is achieved through two primary approaches:

Multi-Agent Orchestration (MAO)

In this method, the system coordinates a set of agents in a predefined sequence, where upstream agents pass data to downstream agents. This structured approach works well when tight coordination is required for specialized tasks.

Multi-Agent Reinforcement Learning (MARL)

Unlike MAO, MARL is a more dynamic approach. Using techniques like Q-learning, the agents operate in a distributed task environment and learn from their interactions with each other and their environment. Q-learning helps each agent optimize its decisions based on past experiences, improving its performance over time through trial and error. MARL allows agents to explore and adapt, optimizing workflows in more complex, unpredictable environments.

Step 5: Achieving Autonomous Execution at Scale

With the support of the LM OS, agentic systems are capable of executing complex workflows, making decisions, and adjusting in real-time. They seamlessly integrate with ITSM systems, update workflows based on new SOPs, and optimize execution through a blend of MAO and MARL.

In environments where real-time decision-making is critical, Q-learning enables the system to refine its strategy, ensuring optimal performance continually. The system adapts to new inputs and conditions, making it a powerful tool for businesses that need to scale operations efficiently and intelligently.

Conclusion: The Promise of Agentic Systems for the Future

Agentic systems are not yet at the forefront of automation, but they show immense potential to transform how businesses manage complex workflows. By combining functions, APIs, LLMs, and UIs into a structured DAG and leveraging the intelligence of an LM OS, these systems can autonomously execute, adapt, and optimize even the most complex business processes.

As agentic systems continue to evolve, integration of Multi-Agent Orchestration, Multi-Agent Reinforcement Learning, and Q-learning will only enhance their ability to scale and optimize workflows. With the right advancements, agentic systems are poised to become a critical tool for businesses seeking to operate more efficiently and autonomously.