Introduction
Our global Emerge series hosted an invitation-only roundtable in New York on September 17th, 2025. Representatives from a variety of industries including financial services, B2B commerce, cloud services, social media technology, and community and social services gathered to discuss “Harnessing the Value of Agentic AI to Drive Competitive Advantage.” The discussion compared agentic AI with programmatic workflows, described a strong agent, and agreed on important points and next steps.
AI agents are goal-seeking loops, not fixed scripts. An agent plans a path to the stated goal, selects and uses tools, inspects results, and then finishes, retries, replans, or asks a human to approve, all within explicit validations and guardrails. The value over traditional automation is adaptability: the path is dynamic while the boundaries (policy, budget, permissions) remain enforced. The conversation drew compelling parallels between two major technological transitions: the evolution from monolithic architectures to microservices, and the current shift from traditional workflows to agentic AI. Where traditional systems rely on rigid, predetermined rules and decision trees, AI agents introduce an intelligence layer that enables context-driven, adaptive decision-making.
This represents a fundamental change in how we approach automation, moving from “if-then” logic to systems that can understand context, adapt to unexpected situations, and make nuanced decisions based on their training and objectives. The discussion highlighted the level of dynamism, automation, and complexity of workflows driven by the use cases and their alignment with business value. This can be driven by goals, data, task complexity, and estimated cost of error.
Core Components Deep Dive
The Four Pillars: Perceive, Plan, Act, Reflect
Our technical discussion centered on the essential components that make AI agents effective:
Perception: How agents gather and interpret information from their environment
Planning: The strategic thinking process for determining optimal actions
Action: Execution through various tools and interfaces
Reflection: Learning and improvement from outcomes
Implementation Strategy Framework
We broke down the concept of agent design into four major sections:
Business Goals: Understanding and having a clear definition of the business goal and, therefore, the associated problem statement, is crucial to the technical design of any system. This becomes even more important when we design dynamic agentic pipelines with multiple components powered by AI intelligence with varying complexity.
Pipeline Design: Once we have a clear understanding of the objective, we break down various components and subcomponents to suit the level of intelligence, automation, and complexity. This helps architect the choice of AI tools and models that constitute the pipeline and, therefore, data, memory, and validation requirements.
Knowledge Design: Once the solution design is finalized, the next key component is designing the knowledge component. This constitutes multiple things: data requirements and representation, how we maintain/store them, and how we utilize them as context augmentation, knowledge grounding, and prompt enhancement.
Checks and Balances: Given the pipeline, we need to specify measurement, metrics, and the mechanism of deploying, monitoring, and utilizing them.
This can be further broken down into specific agent components, which comprise the following:
Inputs and Normalization: This step involves assessing the user touchpoint, the scope, variations, and variety. This highlights the complexity in managing the input and, therefore, the AI component to process, transform, or analyze the same.
Orchestration: Given goal and context, navigate to the appropriate subcomponents whether AI models, task-specific agents, or tools, based on rules or AI-driven decisions.
Executor: Specific AI components would have the following for comprehensive execution of the task:
– Planner/Decomposer: This includes definition of orchestration, success criteria, and stop conditions.
– Memory: This includes task-specific context, knowledge base definition, and its utilization (this means sourcing plus transformation; for example, from structured, unstructured, single or multi-modal and dimensional sources, integration channels, file system, storage or database, simple retrieval, entity-based metadata tagging and filtering, graphical knowledge representation, or knowledge graphs).
– Tooling: Available AI tools, pretrained models, and custom AI components.
– Critic / Evaluator: Layered validators, confidence scoring, and decision briefs for Human in the Loop.
– Observability and Audit: Per-step traces (plan, tool calls, verdicts, outcomes), dashboards/alerts; KPIs drive rollout gates and autonomy levels.
The group conferred on each of these components and how they comprehend this summarized representation with their use cases. The components around solution design resonated with participants around the table with respect to their past experience where they could benefit from pre-evaluation and mapping to business goals and, therefore, align their AI agent design. As triggered during the discussion, the group concluded that some of the key considerations to focus on include safety and policy, reliability, assessment of performance and cost, and evaluation and KPIs. Among all the implementation elements, Executors and Evaluators are two critical aspects of agent design.
Executors emerged as highly flexible components that can be either AI-powered (using LLMs) or programmatic tools, with dynamic selection based on task requirements and context. The key insight was their integration with knowledge bases and memory systems while maintaining workflow execution for expected pathways.
Evaluators were identified as the critical component for system reliability, serving as validation mechanisms for output quality, success criteria evaluation against business goals, and stop condition management to prevent runaway processes. They create continuous feedback loops and integrate with observability systems, essential for enterprise deployment.
Sample Use Cases from Sahaj
Two use cases from Sahaj were presented to showcase practical implementation of the design pipeline discussed. The use cases highlighted AI agents designed with variety in the level of data, orchestration, and goals complexity, and how that transcends into solution design.
The first use case involved development of an autonomous information and knowledge agent for fleet owners of commercial vehicles that impacts their bottom line with efficient resource and machine utilization.
This use case constituted planning around a specific user type and business goal, as stated above. The knowledge requirements spanned across structured, but high-velocity data as well as some unstructured data components. The pipeline design consisted of subagents that decomposed user input, derived information to trigger the knowledge retrieval mechanism (SQL data retrieval using SQL generation agent), and response normalization with user-level account details for personalization.
The SQL generation itself acted as a subagent that had its own data processing and validation components and was retuned to achieve desired accuracy. Overall, system evaluation was designed with a hybrid component of benchmarking, LLM-based critique, and human feedback that helped retune the pipeline.
The second use case had comprehensive business goals that could be further segregated into smaller tasks and, therefore, goals. The main objective was to enhance efficacy in the policy initiative and drafting process for economic mobility by local decision-makers. This meant better information access from knowledge sources created in the form of structured sources like census data, unstructured sources like successful cases, and dynamically created data through user interaction. This transcended into at least three key components around information exploration, decision-assistive recommendation, and finally initial case draft creation.
Given the goals, input, and context, we had a multi-level subagent that targeted, explored, recommended, and drafted. The knowledge design itself utilized an AI tool, GraphRAG, to enhance knowledge representation and also context augmentation. Further submodules performed comprehension of user exploration as well as structured data sources, and were funneled into an AI module to generate assistive recommendations that helped local decision makers (LDMs) plan and decide on policy choices. Once finalized on a policy initiative, the initial step for executing the initiative was assisted by the draft component of the agent. This component incorporated knowledge from shortlisted, bookmarked user interaction with previous explore and recommend components and mapped to standard templates for draft generation.
Overall, the system was able to achieve ease of information access, additional insight and comprehension across information sources, decision-aiding recommendations which were data-backed and assistive in first step execution. The system overall had a validation framework that targeted multifold measures at the task, component, and pipeline levels by using metrics, benchmarks, and human feedback.
Where Agentic Workflows Fit (and Where They Augment)
The discussion transitioned to a generalized view on where agentic workflows benefit significantly as opposed to scenarios where they play a better role as an augmentor rather than a driver.
Agents excel when:
Inputs are messy and context lives across documents and systems.
The work involves support and claims processing, compliance preparation, or cross-system back-office tasks.
The work is knowledge-heavy where the outcome is clear but the path varies by case.
The cases where hard real-time or safety-critical decisions (dosing, avionics), or zero-variance compliance steps where deterministic code is safer, or for already-proven algorithms where plain software is cheaper and provable. In such scenarios, it’s better to use agents to augment (triage, prepare context) draft, but keep the final decision deterministic.
A View Into the Near Future
The near-term trajectory, based on the research direction from AI scientists and companies focusing on foundational model development, includes enhanced reasoning, hybrid AI (SLM + LLM), improved multimodal context, and focus on alignment and safety efficiency. This could further bring in or enhance the current state of AI tools and the level of autonomy and intelligence we associate with them.
Conclusion
Agentic AI is a managerial choice as much as a technical one: define the outcomes and guardrails, wire the loop, and measure relentlessly. The discussion reinforced that successful AI agent implementation requires a thoughtful balance of autonomy and control with careful consideration of where AI adds the most value while maintaining reliability and human oversight where needed.
Some of the key takeaways are:
Architectural Shift: AI agents represent a fundamental change from deterministic workflows to adaptive, intelligence-driven systems.
Component Balance: Success requires careful orchestration of planners, executors, evaluators, and memory systems.
Quality Assurance Evolution: The discussion emphasized that AI agentic applications require more sophisticated quality assurance compared to traditional applications.
Hybrid Strategy: Not everything needs to be “agentic”; strategic automation delivers better outcomes.
Human-AI Collaboration: The most effective implementations thoughtfully integrate human oversight and emotional intelligence.
Continuous Evolution: AI agent systems require ongoing evaluation, tuning, and improvement based on real-world performance.
