Volume 11, Issue 1, January – 2026 International Journal of Innovative Science and Research Technology

ISSN No: -2456-2165 [Link]

Agentic AI Overview
Somil Asthana1
1
Tide. Co

Publication Date: 2026/01/21

Abstract: Agentic AI represents a paradigm shift from traditional single-shot LLM applications to iterative, autonomous
workflows that mirror human problem-solving approaches. Agents range from less autonomous to highly autonomous
enabling spectrum of use.

This article equips data scientists and AI practitioners with the conceptual framework and practical methodology to
design, implement, and optimize agentic AI workflows for complex, autonomous task execution using 4 foundational Agentic
design patterns. Giving pragmatic five-step approach to building agentic workflow.

How to Cite: Somil Asthana (2026) Agentic AI Overview. International Journal of Innovative Science and Research Technology,
11(1), 1344-1349. [Link]

I. INTRODUCTION  Degree of Autonomy

 What Is Agentic AI?  Less Autonomous Agents:

Traditional non-agentic LLM-based solutions, such as
chatbots, operate through single-shot execution—generating  All steps are predefined
outputs directly from prompts. While this approach has  All tools are hard-coded
proven effective for many use cases, it has inherent  Autonomy is limited to text generation
limitations. Agentic AI workflows, by contrast, employ a
multi-step iterative process where agents mirror human  Highly Autonomous Agents:
problem-solving patterns: they analyze, research, revise, and
refine their approach before producing a final result. This  Make many decisions autonomously
iterative loop, built on LLM foundations, may require  Can create new tools or code on the fly
multiple steps and introduce latency, but it consistently  Determine their own processing strategies
delivers superior outputs compared to non-agentic solutions.
 Benefits of Agentic AI Workflows
 Agentic AI Workflow
An Agentic AI workflow consists of multiple ● Effective execution: Agents perform tasks with some
coordinated steps designed to complete complex tasks with accuracy and reliability.
varying degrees of autonomy. These solutions exist along a ● Complex output: Agents can handle sophisticated, multi-
spectrum: faceted tasks that exceed single-shot capabilities.
● Parallelism: Agents can execute multiple intermediate
 Less autonomous agents operate within fully tasks concurrently and synthesize results.
predetermined sequences, executing predefined steps with ● Modularity: Workflows can be modified by replacing
limited flexibility. While effective for their assigned tasks, LLMs or adjusting available tools without restructuring
their capabilities are constrained by their rigid structure. the entire system.
 Highly autonomous agents possess substantial decision-
making freedom, including the ability to determine  Building Blocks for Agentic Workflow
internal processing steps, invoke appropriate tools, A fundamental principle in designing agentic
generate code dynamically, and iteratively refine outputs workflows is decomposition: analyze existing system
through multiple cycles of reflection and improvement. processes and identify discrete steps that can be implemented
These agents handle more sophisticated tasks and produce either through LLMs or specialized software tools. If agents
higher-quality results. cannot execute a step effectively, it should be further
decomposed into manageable substeps.

IJISRT26JAN348 [Link] 1344

Volume 11, Issue 1, January – 2026 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 [Link]

Table 1 Building Blocks for Agentic Workflow

Building Block Example Usecase
Models LLM Text generation, tool use, information extraction
Algorithms Machine Learning Models, Heuristic Models, Prediction, rule-based logic, regression models,
Statistical Regression Models, Python code pdf-to-text, text-to-speech, image analysis
Tools MCP, function tools Web Search, real-time feeds, check social profile,
handle, check calendar, RSS feeds
Information Retrieval Database, Retrieval System like RAG

II. LITERATURE COMPARISON AND The Tool Use pattern described in this tutorial is based
POSITIONING on the ReAct (Reasoning and Acting) framework introduced
by Yao et al. (2023) transforms the idea into practical
This tutorial aligns closely with foundational work in implementation.
agentic AI while providing a practical, implementation-
focused perspective for data scientists and practitioners. The Our evaluation framework uses objective (Python-
four design patterns presented in this tutorial—Reflection, based) and subjective (LLM-as-judge) approaches as
Tool Use, Planning, and Multi-Agent—align directly with the documented by (Raschka, 2024; Chang et al., 2024)
framework popularized by Andrew Ng in his 2024
presentations and [Link] courses on agentic  Agentic Design Patterns
workflows (Ng, 2024). This tutorial extends his framework Four critical agentic design patterns form the
by providing detailed implementation guidance, evaluation foundation of sophisticated workflows. These patterns can be
strategies, and practical workflows for production combined in various configurations to create complex,
deployment. powerful systems.

Table 2 Agentic Design Patterns

Design Pattern Design Pattern Description
Reflection An LLM (the same or a different one) reviews previously generated text, evaluates its quality, and suggests
improvements. This iterative critique process enhances output quality.
Tool use LLMs are provided with tools (functions) to assist with assigned tasks. Examples include code execution for
analytical tasks, database resources and web search for information gathering, email/calendar tools for
productivity, and image libraries for visual tasks.
Planning A collaborative workflow incorporating four stages: (1) Task Planning, (2) Model Selection, (3) Task
Execution, and (4) Response Generation. This framework enables flexible, autonomous
[Link].
Multi-agentic Multiple agents coordinate, debate, and converge toward consensus. While challenging to control,
workflow preliminary research indicates these workflows can achieve exceptional accuracy.

 Agentic Pattern: Reflection identifying inconsistencies and bugs. Incorporating external

Just as humans reflect on their work to identify feedback into the reflection process further enhances
improvements, LLMs can be prompted to evaluate their own generation efficiency and quality.
outputs. Reflection is particularly effective in code
generation, where reasoning-focused LLM models excel at The below figure illustrates a reflection pattern flow.

Fig 1 Reflection Design Pattern

IJISRT26JAN348 [Link] 1345

Volume 11, Issue 1, January – 2026 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 [Link]

 Agentic Pattern: Tools autonomously decides whether and when to invoke specific
Tools extend LLM capabilities beyond text generation, functions. After execution, function outputs are fed back into
enabling agents to take actions, retrieve information, and the conversation history, allowing agents to generate
perform computations. In this pattern, tool information is informed final responses.
provided to the agentic workflow, and the LLM

Fig 2 Agentic Tools Design Pattern

Table 3 Agentic Pattern: Tools

Scenario Tool Expected Output
Prompt asking to search for nearby web search tool with location and date Agents combine tools to provide
recreational activity. time tool information about clubs or activity centers
Business Intelligence asking for specific SQL query creation tool and SQL Agents generate SQL queries, execute
query on data execution Python function them via database connections, and
provide business intelligence insights
Complex calculations to compute ROI Function call to retrieve asset prices Agents calculate current portfolio ROI
for a portfolio combined with ROI calculation tool

Modern LLMs no longer require extensive descriptions functionalities as callable functions, dramatically expanding
in prompts to invoke functions. Today's models are their power and reach.
increasingly trained to interpret and call functions based
solely on function names and docstrings, demonstrating  Agentic Pattern: Planning
enhanced logical decision-making capabilities. Agentic workflows become significantly more flexible
when planning components allow for dynamic rather than
One of the most powerful tools available to agents is fixed execution flows. In this pattern, LLMs determine the
code execution, which allows agents to generate Python code optimal course of action rather than following predetermined
that executes in the background. This transforms agents from sequences. A planner LLM receives functional tools and
simple text generators into computational engines capable of prompts as inputs, it generates step-by-step instructions
sophisticated planning and execution. (plan) required for achieving desired outputs. Each
instruction/task, along with relevant context, prompt and
Another significant advancement is the Model Context tools, is passed to subsequent LLMs to run, forming a
Protocol (MCP), which extends function tool capabilities to processing chain from input to final output.
remote servers, effectively allowing agents to leverage server

IJISRT26JAN348 [Link] 1346

Volume 11, Issue 1, January – 2026 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 [Link]

Fig 3 Agentic Planner Design Pattern

This pattern eliminates the need to predefine function III. EVALUATION

call sequences. Instead, agents use available tools and
prompts to formulate plans—typically in JSON format— Building an optimal agentic workflow is inherently
specifying which function tools subsequent LLMs should iterative—it's difficult to know in advance which components
invoke. While this approach reduces control over execution will succeed and how to evaluate them effectively. A best
flow, it significantly enhances adaptability and problem- practice is to build a "quick and dirty" initial prototype,
solving capability. examine its outputs (including errors and unsatisfactory
results), and then systematically improve it.
 Agentic Pattern: Multi-Agent
While it might seem that extending a single agentic Start by collecting 10-20 representative inputs and
workflow with additional function tools or layers would running them through the agentic workflow to identify
suffice, multi-agent patterns offer distinct advantages. This common output errors. This empirical approach quickly
approach enables agent reuse across domains and promotes reveals the most frequent failure modes. For more systematic
modular development. For example, an agentic workflow evaluation, develop automated evaluation functions to track
designed for creating graphic visualizations can be reused specific error types. For instance, in an invoice processing
across multiple applications. Additionally, teams can develop workflow, testing 10-20 invoices might reveal that date
agentic solutions independently and in parallel, then integrate extraction is the most common error. You can then create an
them later. evaluation function specifically for date extraction, measure
accuracy, implement improvements, and monitor changes
Multi-agent communication remains an active research over time.
area, with ongoing work on optimal communication patterns
and coordination protocols. Despite control challenges, this Errors fall into two categories: objective and subjective.
pattern offers substantial flexibility and scalability benefits. Objective errors involving numerical outputs can be
evaluated using Python functions. Subjective errors,
particularly those involving text quality, benefit from an
"LLM-as-judge" approach, where outputs are compared
against 3-5 gold standard examples for each domain.

Table 4 Two “Axes” of Evaluation

Evaluate with python code LLM-as-judge (subjective)
Target variable is Checking for certain numerical variables like invoice date. Count number of gold standard points
number or count Python Eval Function generated by agentic workflow
Target variable is Checking for text length Decoding images and charts
text

IJISRT26JAN348 [Link] 1347

Volume 11, Issue 1, January – 2026 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 [Link]

IV. STEPS FOR CREATING AN AGENTIC  Examples of SLMs:

WORKFLOW
 Text generation: Llama 3.2-1B, DeepSeek-R1-1.5B, Phi-
The challenge in creating agentic workflows lies in the 3.5-Mini-3.8B
multitude of possible approaches: Should you invest time  Multi-modal: Phi-4-multimodal, Gemma 3-4B
perfecting workflow design or focus on LLM selection? With  Code generation: CodeGemma (Google), StarCoder2
numerous design patterns available, the process can seem (BigCode/HuggingFace)
daunting.
 Step 4: Implement a Workflow
 Step 1: Plan a Quick and Direct Workflow Prototype Write code and develop prompts to connect LLMs and
Avoid spending weeks or months theorizing about an construct an initial prototype.
ideal solution. Given the novelty of these systems, it's more
effective to create a rapid, imperfect prototype that addresses  Step 5: Conduct an Error Analysis
even a subset of requirements. This provides a concrete Error analysis identifies underperforming components
foundation for iterative improvement. that degrade overall system performance. This process helps
prioritize improvements such as: one might need to replace
 Step 2: Select or Design Functional Tools an LLM with a more capable model or, revise tools
Before evaluating LLM and trade-offs, identify and (computational functions, web search, MCP) that produce
implement the function tools agents will need. For an invoice substandard outputs, or refine prompts with more explicit
processing workflow, this might include: (1) PDF-to-text instructions and examples.
conversion, (2) date extraction tool, and (3) database update
functionality for storing invoice details. Focus on the component producing the least satisfactory
intermediate output. Conduct component-level error analysis,
 Step 3: Initially Choose Small Language Models (SLM) or implement improvements, then move to the next problematic
Larger SLMs component. This systematic approach drives overall
Selecting the appropriate LLM—balancing general- performance improvement and reveals workflow bottlenecks.
purpose versus specialized multi-modal capabilities—can be
complex. For initial prototyping, start with Small Language  Example: Stock Portfolio Agentic Workflow
Models (SLMs), which can be replaced with more powerful An example workflow for stock portfolio management
LLMs as needed. connects to a broker account to fetch current information,
performs profit and loss calculations, and summarizes results
in a database.

Fig 4 Example: Stock Portfolio Agentic Workflow

IJISRT26JAN348 [Link] 1348

Volume 11, Issue 1, January – 2026 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 [Link]

 One Needs Following Tools International Conference on Learning

Representations (ICLR).
 Broker MCP get_holdings [Link]
 Broker MCP place_order
 Broker MCP cancel_order
 Broker MCP get_quotes
 Broker MCP get_ltp
 Broker MCP get_historical_data
 Web tool
 Calculate_Sharpe_Ratio
 Calculate_Sortino_Ratio
 Calculate_Stock_PL
 Calculate_Portfolio_PL
 Update_database_stock_pl
 Update_database_portfolio_pl

 LLMs to Complete Following Tasks

 LLM to decode the user query extract broker account

details
 LLM to extract assets from the user's portfolio, identify
the symbols. Fetch their LTP.
 LLM to summarize PL statements along with individual
stock performance.

Finally, prompts along with context to stitch a complete

workflow.

V. CONCLUSION

Our novel contribution in writing this article 1)

Integrated Methodology: Rather than treating design patterns,
evaluation, and implementation in silo, we present an
integrated five-step guide for practitioners to prototype
Agentic workflow to production. 2) Emphasis on Rapid
Prototyping: Agentic AI is still an active area of research and
developers are trying to figure out best way to design and
code therefore, a "quick and dirty" prototype-first approach
followed by evaluation can assist in developing a workable
Agentic AI workflow that can be improved as one monitors
and verify the output. 3) Small Language Model (SLM)
Focus: Our recommendation to start with SLMs (Llama 3.2-
1B, DeepSeek-R1-1.5B, Phi-3.5-Mini) for initial prototyping
is often overlooked in practice - engineers and businesses
want to the best LLM and tools fail at times because of the
complexity involved in designing an Agentic workflow.

REFERENCES

[1]. Ng, A. (2024). Four design patterns for AI agentic

workflows. LinkedIn Post, March 2024. Retrieved
from
[Link]
agent-for-many-worlds-cross-species-activity-
7179159130325078016
[2]. Ng, A. (2024). Agentic AI [Online Course].
[Link]. Retrieved from
[Link]
[3]. Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I.,
Narasimhan, K., & Cao, Y. (2023). ReAct:
Synergizing reasoning and acting in language models.

IJISRT26JAN348 [Link] 1349