What Is Agentic AI?

Agentic AI refers to artificial intelligence systems that can autonomously plan, decide, and act to achieve a goal - without requiring step-by-step human instructions. Unlike traditional AI that responds to a single prompt with a single answer, an agentic system breaks down complex tasks, uses tools, adapts to feedback, and orchestrates multi-step workflows on its own.

The shift from "AI that answers" to "AI that acts" became a central theme of enterprise AI in 2025 and 2026. Chatbots answer questions. Agents do work - they query databases, update records, generate reports, trigger workflows, and escalate when they encounter something outside their scope. This makes data governance not just important for agentic AI, but a prerequisite: agents that act on ungoverned data can cause damage at machine speed.

What Makes AI Agentic

Not every AI system that uses a large language model (LLM) is agentic. The distinction lies in the degree of autonomy the system has in deciding what to do, how to do it, and when to stop.

A traditional LLM interaction is stateless: the user sends a prompt, the model generates a response, the interaction ends. An agentic system adds three capabilities on top of the language model:

• Planning - the agent decomposes a high-level goal into subtasks and decides the order to execute them. It doesn't just respond to instructions - it generates its own plan of action.
• Tool use - the agent can invoke external tools: databases, APIs, code interpreters, web search, file systems. This grounds the agent in real data and gives it the ability to take actions beyond text generation.
• Autonomous iteration - the agent evaluates the results of its actions, decides whether the goal has been met, and adjusts its approach if not. This reasoning loop continues until the task is complete or the agent determines it needs human input.

The combination of these three capabilities is what transforms a language model from a sophisticated autocomplete into a system that can independently accomplish work.

How AI Agents Work

At a fundamental level, an AI agent operates through a reasoning-action loop - sometimes called the "observe-plan-act-evaluate" cycle. The agent receives a goal, observes the current state of the world (through tool calls or provided context), plans the next step, executes it, evaluates the result, and repeats until the goal is achieved.

The LLM as the Reasoning Engine

The large language model at the core of an agent serves as the reasoning and planning engine. It interprets the goal, decides which tools to use, generates the parameters for tool calls, and evaluates results. The LLM doesn't "think" in the human sense - it generates a plan by predicting the most likely useful sequence of actions based on its training data and the current context.

Agent frameworks implement this loop in different ways. LangGraph describes itself as a low-level orchestration framework for long-running, stateful agents and models a workflow as a graph of steps over shared state. OpenAI released its Agents SDK in March 2025, and Microsoft's AutoGen (now in maintenance mode, succeeded by Microsoft Agent Framework), CrewAI and Anthropic's Claude Agent SDK offer other abstractions for the same pattern. In each, a step produces an observation that updates the agent's context, and the LLM decides the next action based on the accumulated state. This allows the agent to adapt mid-execution: if a database query returns unexpected results, the agent can reformulate its approach rather than failing.

Data platforms also offer managed services for building agents next to the data, such as Databricks Agent Bricks, Snowflake Cortex Agents and Microsoft Foundry Agent Service.

Tool Integration: MCP and Function Calling

Tools are what give agents their capabilities. Without tools, an agent is just an LLM generating text. With tools, it can query databases, call APIs, read documents, execute code, and update systems.

Two primary mechanisms enable tool integration:

• Function calling - the LLM generates a structured JSON call to a predefined function, which the application executes and returns results to the agent. This is the foundational mechanism, supported by the major model providers, including OpenAI, Anthropic and Google.
• Model Context Protocol (MCP) - an open protocol that Anthropic introduced on November 25, 2024 to standardize how AI applications connect to tools and data sources. Instead of building custom integrations for each tool, MCP defines a standard protocol for tool discovery, invocation, and resource access. In December 2025 Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation. Teams adopt it because it separates the agent logic from the tool implementation.

The choice of tools available to an agent directly determines what it can do - and what it can access. This is where data governance becomes critical: an agent with access to ungoverned data can produce confidently wrong results, or worse, take autonomous actions based on incorrect information.

Memory and Context

Agentic systems need to maintain state across the reasoning loop. This happens at multiple levels:

• Short-term memory - the conversation context and accumulated observations within a single task execution. This is typically the LLM's context window, augmented by the tool results collected during the session.
• Long-term memory - persistent storage that allows the agent to recall information from previous interactions. Implemented through vector databases, knowledge graphs, or structured storage systems.
• Shared memory - in multi-agent systems, a common workspace where agents deposit and retrieve information for coordinated work.

Agentic Patterns

Not all agentic systems look the same. The level of autonomy and complexity varies significantly based on the use case. These patterns represent increasing levels of agentic behavior.

Single-Agent Task Execution

The simplest agentic pattern: one agent, one goal, a set of tools. The agent receives a task (e.g., "analyze last quarter's sales data and summarize trends"), plans the steps, executes tool calls (query database, process data, generate summary), and returns the result. This pattern is suitable for well-defined tasks with clear completion criteria.

Tool-Augmented Reasoning

The agent uses tools not just to take actions, but to enhance its reasoning process. For example, an agent tasked with answering a complex question might first search a knowledge base, then query a database for supporting data, then cross-reference the results before generating an answer. Each tool call informs the next reasoning step. This pattern is common in retrieval-augmented generation (RAG) systems.

Multi-Agent Orchestration

Complex workflows are decomposed across multiple specialized agents, each with their own tools and expertise. A coordinator agent (or an orchestration framework) manages the workflow: assigning subtasks, collecting results, handling dependencies, and synthesizing the final output. For example, a data quality review might involve a profiling agent, a rules-checking agent, and a reporting agent working in sequence.

Agents built on different frameworks need a shared protocol to work together. Google introduced the Agent2Agent (A2A) protocol for this in April 2025, and the Linux Foundation began hosting it as a vendor-neutral project in June 2025.

Autonomous Workflows

The most advanced pattern: agents that operate continuously, monitoring conditions and taking actions when triggers are met. An autonomous data stewardship agent might monitor data quality metrics, flag anomalies, generate incident reports, and notify the responsible data steward - all without human initiation. This pattern requires the strongest governance guardrails because the agent operates without per-action human approval.

Data Governance for Agentic AI

Agentic AI doesn't just consume data - it acts on data autonomously. This changes the governance equation fundamentally. When a human analyst queries a database, they apply judgment about data quality, relevance, and context. When an agent queries the same database, it lacks that institutional knowledge unless governance infrastructure provides it.

The stakes are practical. In June 2025 Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear business value or inadequate risk controls.

Why Agents Need Governed Data

An agent making autonomous decisions based on data is only as trustworthy as the data it accesses. Three governance capabilities become critical:

• Data quality signals - agents need machine-readable indicators of data quality (freshness, completeness, accuracy scores) to decide whether a dataset is reliable enough for the task at hand. Without quality metadata, agents treat all data as equally trustworthy.
• Business context - an agent querying a field called margin needs to know whether it means gross margin, net margin, or contribution margin. Business glossary terms and semantic metadata provide the context that prevents agents from making confidently wrong interpretations.
• Access governance - agents should only access data they are authorized to use for their specific purpose. Role-based access controls, data classification, and purpose-bound access policies prevent agents from reading sensitive data they don't need.

Regulation and Security Frameworks

The EU AI Act sets risk-based requirements, and the strictest apply to high-risk AI systems: those listed in Annex III, such as AI used in recruitment or creditworthiness assessment, and AI that serves as a safety component of products covered by the legislation in Annex I. Whether an agent is high-risk depends on what it is used for, not on its architecture. High-risk systems must meet requirements that include data governance, transparency, human oversight, and accuracy:

• Article 10 requires training, validation and testing data sets to follow data governance and management practices and to meet quality criteria, which depends on data governance infrastructure.
• Article 14 requires high-risk systems to be designed so that people can effectively oversee them while they are in use, which for agents usually means human-in-the-loop checkpoints for high-stakes actions.
• Article 13 requires enough transparency for deployers to interpret a system's output and use it appropriately, with instructions for use; data lineage and audit trails help show which data an agent relied on.

The dates have moved. Obligations for general-purpose AI models have applied since August 2, 2025. Regulation (EU) 2026/1744, the AI Omnibus, moved the high-risk obligations to December 2, 2027 for Annex III systems and to August 2, 2028 for AI in Annex I products.

Outside the EU, the NIST AI Risk Management Framework (AI RMF 1.0, January 2023) offers a voluntary structure for managing AI risk. For security, OWASP's Top 10 for LLM Applications 2025 lists excessive agency (LLM06) as the risk of an LLM-based system taking damaging actions on unexpected, ambiguous or manipulated model output, and in December 2025 OWASP published a separate Top 10 for Agentic Applications.

How Dawiso Enables Agentic AI

Dawiso supplies the governed context that agentic AI systems depend on; it does not run or orchestrate agents itself. Through the Dawiso MCP server, any MCP-compatible agent can work with Dawiso's catalog, glossary, and lineage over the Model Context Protocol. The documented tools cover search, object details, lineage and data samples, plus write tools such as creating objects and updating attributes or workflow states, and the tools an agent sees follow the connecting user's role in Dawiso.

What this means in practice: an AI agent tasked with analyzing business data can first search Dawiso's catalog to find relevant datasets, look up business glossary definitions for field-level context, check ownership to know whom to ask, and trace data lineage to understand where data comes from, all through MCP tool calls, before it queries the actual data source. Dawiso does not measure data freshness or run quality checks itself: quality checks are documented in Dawiso together with the external engine that runs them, so live quality signals still come from those tools.

The Business Glossary gives agents shared definitions, and its terms are linked automatically wherever they appear in object content. The Data Catalog makes the metadata that connectors scan discoverable. Interactive Data Lineage lets agents trace data provenance, with lineage built from four sources: connectors, dbt, SQL parsing (not switched on by default) and relations that users add by hand. And when AI-powered features are enabled, Dawiso generates descriptions for databases, schemas, views and tables that stay marked for review until a person saves them.

Conclusion

Agentic AI represents a fundamental shift from AI that answers questions to AI that autonomously accomplishes work. The reasoning loop - observe, plan, act, evaluate - combined with tool integration gives agents capabilities that go far beyond text generation. But this autonomy creates a critical dependency on data governance: agents that act on ungoverned data can cause harm at scale.

For organizations adopting agentic AI, the message is clear: invest in data governance before you invest in agent capabilities. A well-governed data landscape - with cataloged assets, quality metrics, business glossary terms, and proper access controls - is the foundation that makes agentic AI trustworthy. Without it, you're giving autonomous systems the keys to a building where nobody labeled the rooms.

Sources

• Anthropic, Introducing the Model Context Protocol, November 25, 2024.
• Anthropic, Donating the Model Context Protocol and establishing the Agentic AI Foundation, December 9, 2025.
• Google Developers Blog, Announcing the Agent2Agent Protocol (A2A), April 9, 2025.
• Linux Foundation, Linux Foundation Launches the Agent2Agent Protocol Project, June 23, 2025.
• OpenAI, New tools for building agents, March 11, 2025.
• LangChain, LangGraph overview.
• Databricks, Agent Bricks.
• Snowflake, Cortex Agents.
• Microsoft Learn, What is Microsoft Foundry Agent Service?
• Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, press release, June 25, 2025.
• EUR-Lex, Regulation (EU) 2024/1689 (Artificial Intelligence Act).
• EUR-Lex, Regulation (EU) 2026/1744 (AI Omnibus).
• NIST, AI Risk Management Framework.
• OWASP, LLM06:2025 Excessive Agency.
• OWASP, OWASP Top 10 for Agentic Applications.
• Dawiso Help, Dawiso Model Context Protocol (MCP).
• Dawiso Help, Data product quality and SLAs.
• Dawiso Help, Business glossary.
• Dawiso Help, Sources of data lineage.
• Dawiso Help, AI features and integrations in Dawiso (AI features).