Why what an agent knows matters as much as what it can do

This blog is part of a series.

Read previous part: Designing the Agent Control Plane

Read next part: Small and Domain-Specific Language Models

Agentic systems combine behaviour and knowledge. A system can behave correctly and still produce the wrong outcome if the information it relies on is incomplete, outdated, or incorrect.

An agent that interprets its goals correctly, follows its plan diligently, and stays within its permitted boundaries can still cause significant problems if the information it is acting on is stale, incomplete, or wrong. This is one of the less intuitive aspects of agentic systems for organisations encountering them for the first time. The instinct is to focus on what the agent does. The Data Plane is concerned with what the agent knows, and how it knows it.

Getting this right is not optional. It is the difference between an agent that is genuinely useful and one that is fluent but unreliable.  The Data Plane is the layer that governs how an agent accesses, interprets, validates, and maintains information throughout a task.

The problem with model knowledge alone

Large language models are trained on vast amounts of data. They have broad knowledge across many domains and can reason about complex topics with apparent sophistication.

This creates a dangerous illusion: that a well-prompted model is a well-informed agent.  It is not.

Model knowledge has a training cutoff. It reflects the world as it was, not as it is. It contains biases and gaps that are difficult to detect from the outside. And critically, it contains no knowledge of your organisation. It does not know your systems, your data, your rules, or your context.

An agent operating on model knowledge alone will fill those gaps with plausible-sounding assumptions. It will answer questions about your customer records without having seen them, make decisions about your workflows without understanding them, and act on its best guess about your business rules rather than the actual rules.  The outputs will often look reasonable. They will sometimes be wrong in ways that matter.  Good behaviour on bad information is still failure.

The Data Plane is the architectural response to this problem. It gives agents access to real, current, and validated information rather than relying on what the model assumes to be true.

What the Data Plane provides

The Data Plane ensures that the agent is working with information that is current, relevant, and reliable. It does this through a set of components that operate across the lifecycle of a task.

• Retrieval and knowledge access
  These are the mechanisms by which an agent accesses information from outside the model. This includes vector databases for semantic search, structured data stores for precise lookups, document repositories, and API connections to live systems. The design of these systems determines the quality of the information the agent receives. Poor retrieval returns irrelevant or outdated content. Well-designed retrieval provides the right information at the right time.
• GroundingGrounding anchors agent outputs to verified sources rather than allowing the model to generate from its own internal knowledge. A grounded agent cites sources, checks assertions against retrieved content, and flags uncertainty when reliable information cannot be found. Grounding is what separates an agent that is reliable from one that merely sounds reliable.
• State management and memory
  Agentic tasks often span many steps. The agent must maintain context across the entire sequence. This includes what has been retrieved, what has been decided, and what has already been executed. Without structured state management, agents repeat work, lose context, and make inconsistent decisions. Memory defines what is retained, for how long, and in what form.
• Structured validation and domain rules
  These mechanisms check agent reasoning and outputs against the rules of the domain. A domain rule might require specific evidence for a decision, enforce valid ranges for outputs, or define conditions that must be met before an action is allowed. These rules encode organisational knowledge in a form that can be applied consistently.

The role of specialised models

General-purpose language models are capable, but they are not always the right tool for every task within an agentic system.

Specialised models play an important role within the Data Plane. A model trained on medical data will outperform a general model on medical classification tasks. A model trained on financial data will apply domain rules more reliably than one reasoning from general knowledge.

These models are not replacements for the orchestrating model. They are components within a broader pipeline, invoked at specific points to perform tasks with greater accuracy and efficiency.

In production systems, it is increasingly common to use multi-model pipelines, where different models handle different aspects of reasoning and validation. The Data Plane is where those pipelines are designed and governed.

What breaks without a Data Plane

When the Data Plane is poorly designed, the problems are predictable.  Agents act on stale information because retrieval is not connected to live systems. They make inconsistent decisions because state is not maintained across a long task. They produce outputs that contradict organisational rules because validation is absent. They hallucinate details because grounding is incomplete and gaps are filled with assumptions.

These failures are difficult to detect from the output alone. A response can be well-structured, confident, and plausible while still being wrong.  The issue is not how the output looks. It is whether it reflects reality. Failures often surface only when outputs are checked against real data or when their downstream consequences become visible. An agent whose behaviour is well-governed, but whose information is unreliable is still unreliable.

What good looks like

A well-designed Data Plane ensures that an agent always knows where its information came from, how current it is, and how confident it should be in it.

Retrieval is precise and purposeful. State is maintained consistently across the full task. Domain rules are applied at the right points. When reliable information is not available, the agent does not guess. It says so. For business leaders, the Data Plane answers a simple but critical question: how do we know the agent is working with the right information?

The answer is about architecture, not just prompting. It requires deliberate design of retrieval, memory, validation, and domain knowledge. And it requires ongoing maintenance as the information landscape changes.

An agent that acts well on good information is a valuable asset. An agent that acts well on bad information is a liability that is difficult to spot until the damage is done.

Learn more about Bell’s AI Solutions.

Meet the Author:

Alison Hodivala
Director of AI & Data Portfolio Strategy at Bell Integration

Alison Hodivala is the Director of AI & Data Portfolio Strategy at Bell Integration, where she focuses on ensuring the company’s AI and data services reflect market needs and are supported by delivery models that scale reliably. She works across product, delivery and commercial teams to develop offerings that are practical, relevant and capable of delivering business impact.

Her career spans more than 25 years in start-ups and global organisations, covering consultative sales and delivery, enablement and change management. Drawing on a background in organisational psychology, Alison applies an evidence-based approach to AI adoption, aligning people, processes and technology to achieve effective and responsible outcomes.