Software agents are only as smart and capable as their design
Agent ABCs
Agents are simply software entities that execute tasks on a user’s behalf. Beyond a typical “bot”, they are expected to perform complex automation tasks and use context to optimize process outcomes.
Agents are proxies for human assistants that handle routine, tedious or time-consuming work as a background process. They allow people to delegate work so they can take on more responsibilities.
Technically, agents are algorithms exposed by an interface designed to receive high-level instructions and then independently perform related actions with little or no manual intervention. They are capable of constructing and executing workflows as well as integrating other relevant information sources and tools.
Processing generally starts with agents receiving instructions, high-level commands, from a customer, a business user or developer. Agents then translate the user’s intent and decide on a course of action. They break the work down into an ordered set of steps as well as related operations, which may include API calls to fetch relevant data and trigger required functions. The agents execute the process to fulfill the original command.
Agents are part of the evolution of software development from statically coded and manually integrated solutions that were not intuitive, responsive or agile to dynamic and highly-interactive human-computer interaction that is easy-to-use, contextual and can evolve with changes in the domain or environment. While seemingly new, these concepts of artificial intelligence and software agents goes back to the 1950s.
Like their human counterparts, agents come in many types, shapes and sizes with varying degrees of autonomy and reasoning power.
In general, and at a high-level, all agents require:
1} an interface to receive instructions and support interactions with humans or other agents
2) long and short-term memory for processing multi-step workflow and persisting logs and state changes
3) API access to connect with the environment including related business objects
4) methods for reasoning (e.g., interpreting user intent and conditions) and planning (e.g., deciding which tasks to include, what data sources to query and tools to call, in what order)
The scope and complexity of agent decision-making and actions varies by design. Agents can range from rules-based and platform dependent with a graphic User Interface to AI-driven and truly autonomous standalone entities with Natural Language Programming (NLP) interfaces. As with software architecture generally, there are trade-offs for design decisions (e.g., accuracy, consistency, security, performance, scalability, cost and energy efficiency).
As organizations consider agent-based solutions for a wide-range of automation use cases and more broadly for transformation initiatives, it is important to cut through the hype and objectively match agent capabilities to functional requirements.
The new agents on the block
Agentic AI emerged only recently, Spring of 2023, as a byproduct of the explosion of interest in generative AI. There was a near immediate interest to leverage GenAI for automation, which resulted in Agentic AI - agent-based solutions that leverage LLMs for reasoning and planning.
In the race for Agentic AI, the easiest launching point was Robotic Process Automation (RPA). RPA tools use bots for scraping content from internal and public websites to provide an informal type of data integration. RPA products embraced simplicity, leveraging relatively basic If-This-Then-That (IFTTT) workflow logic. IFTTT is a popular approach to simple event-driven workflows that is widely used for integrating web and social media apps.
Simple is as simple does
Early Agentic AI implementations are a bit of an odd duck in that they generally combine state-of-the-art Neural Networks with fairly primitive agent design (e.g., python scripts, FSMs, etc.). Moreover, since Agentic AI is dependent on LLMs, it inherits generative AI’s strengths and weaknesses (i.e., inaccuracy, inconsistency and hallucinations, as well as security, latency, resource and energy consumption issues). As a result, Agentic AI tends to have limited reasoning and planning capabilities,
Efforts to improve Agentic AI generally circle around improving LLM outputs with techniques like prompt engineering, retrieval augmented generation (RAG), Chain-of-Thoughts and Mixture of Experts. However, doubling-down on LLMs only marginally improves GenAI output accuracy and explainability, and isn’t able to eliminate hallucinations. The LLM optimization techniques are workarounds that add cost, complexity and latency to Agentic AI solutions. No matter how much is spent on training, tuning and tokens, LLMs never becomes deterministic (i.e., accurate, repeatable, explainable, transactional).
Bolt-on agents extend old systems
In parallel to Agentic AI, established software vendors started to adopt GenAI NLP interfaces and integrate agents with their existing products. These AI-enabled solutions generally offer more mature integration and workflow capabilities.
Enterprise-grade, agent-based automation
EnterpriseWeb bridges the gap between last year’s no-code platforms and this year’s agentic AI. Rather than simply bolt-on agents and LLMs to an existing product offering, EnterpriseWeb was built from the ground up for agent-based, ontology-driven, contextual automation at scale. The company holds 22 awarded US patents. Additional patents pending.
EnterpiseWeb is the first agent-based automation platform that supports complex, enterprise-grade processes. Developers can “talk to the platform” to design, deploy and manage intelligent business and infrastructure applications with security, IT governance and AIOps.
EnterpriseWeb’s agents are virtual software engineers that support the end-to-end service design lifecycle from onboarding and composing solution elements, to generating deployment workflows and instantiating observability.
Behind the scenes, agents autonomously call the tools they need for communication, integration, configuration and orchestration tasks. EnterpriseWeb includes a library of serverless middleware functions so the agents can efficiently process requests locally (i.e., no third-party queues, gateways, brokers, integration tools, workflow engines or orchestrators required). Agents can also orchestrate external systems, services, databases and devices for end-to-end automation.
EnterpriseWeb Agents dynamically assemble and configure tasks into complex, multi-step workflows and long-running processes. Instead of a static flowchart or pipeline, agents construct the process step-by-step based on conditions. Agents can delegate tasks to other agents, which perform them in parallel, improving performance.
Knowledge is power
In contrast to LLM-centric Agentic approaches, EnterpriseWeb agents leverage a rich graph model (i.e., an ontology) that enables them to deterministically interpret events, evaluate conditions, automate decisions and optimize actions. Agents use domain knowledge and real-time operational context from the graph, along with inferences from AI and analytics, to personalize user experiences, optimize transactions and synchronize operations.
EnterpriseWeb’s ontology gives agents a logical model with shared domain semantics, metadata and objects for fast access to facts as well as access to distributed solution elements making the agents situationally aware. The design is highly-reflective and deeply recursive for advanced reasoning and planning capabilities.
EnterpriseWeb’s graph is the central source of context, a single-source-of-truth for agent processing. The rich graph eliminates the need for model training and minimizes tuning and tokens. It allows for highly targeted Large Language Model (LLM) interactions that greatly reduce LLM resource and energy costs, while improving LLM outputs and inferences.
Accelerating time-to-value
The lightweight (~50MB), cloud-native platform deploys in minutes and is ready-to-use. It can run in the cloud, on-premise or at the edge.
The agents have a generative AI powered natural language interface, as well as a traditional graphic user interface (GUI) and application programming interfaces (APIs), which are all synchronized so developers can flexibly switch between them. This allows the agent platform to act as a backend service for a customer’s existing portal.
EnterpriseWeb is a transformational agent-based platform that speeds and eases the adoption of advanced AI automation capabilities. It improves developer productivity and enables intelligent and agile operations that optimize business outcomes.
Copyright 2024, EnterpriseWeb LLC
Related Links:
BLOG: Agentic Automation: Knowledge is Power
Demo: EnterpriseWeb: Telco-grade GenAI for intent-based Orchestration
Demo: EnterpriseWeb: GenAI for Business ETL and Code Support 062624
Demo: EnterpriseWeb: Secure Dev Centric Networking CAMARA APIs and Intel TDX
