AI Agent Basics: 6 Components Driving Agent Performance

Summary

Strong AI agent performance depends less on the model alone and more on the system built around it. In this blog, John Briggs, CEO of Screendragon, explains the basics of AI agents and details the 6 components that enable marketing teams to drive agent performance.

For marketing teams and agencies, the best agents are supported by structured knowledge, approved assets, connected tools, validation loops, clean context, memory, and cost controls.

With the recent release of Screendragon’s AI Hub, the platform now enables enterprise marketing teams and agencies to build, deploy, and govern their own AI agents directly inside the workflows that already power their business, so they can harness AI at scale without losing control.  

This functionality is essential to organisations as it determines whether AI agents can operate safely, consistently, and cost-effectively inside marketing workflows, rather than simply generating one-off outputs in isolation.

1. Context Management and Memory

Context management is one of the most underrated factors in AI agent performance. 

An AI model can only work with the information available in its active context window. This is the working space the model uses to understand the task, follow instructions, and produce an output. What goes into that space matters enormously. 

In many cases, the quality of the output depends less on which model is being used and more on what information the model has been given. A strong model with messy context can still produce weak results. A well-designed agent with clean, relevant context is far more likely to produce useful, accurate, and consistent work. 

Researchers have called this problem “context rot”. Most people have experienced this without realizing it. The longer you talk to an AI, the more it can sometimes feel like it is getting “dumber.” It forgets the original goal. The model is not losing intelligence. It is working from a messy context. 

Effective context engineering is about deciding what the agent should pay attention to at any given moment and controlling when information is introduced. 

Techniques such as conversation summarisation, observation masking, and progressive disclosure can all improve output quality and reduce cost. A better agent keeps the context clean. It knows what to use now, what to retrieve later, what to summarise, and what to ignore. 

That is what allows agents to stay focused as work becomes more complex. 

2. Knowledge Bases and Retrieval (RAG)

Rather than injecting an entire brand guidelines document into every content generation call, a well-structured knowledge base retrieves only the relevant sections at the right moment. 

Why is this important? Because AI agents have a limited amount of information they can work with at one time. Every document, instruction or example added to a prompt uses part of that available space. In AI terms, this is measured in tokens. So, if a 50,000-token brand guidelines document is included every time an agent writes a piece of content, the model has to process a large amount of information that may not be relevant to the task. 

That creates two problems. First, it increases cost because the model is processing more input than necessary. Second, it can reduce performance because the agent has to sort through too much context before deciding what matters for its output. 

A better approach is selective retrieval. For example, if the agent is asked to write in the brand’s voice, it should retrieve the two paragraphs that explain tone, not the entire brand document. 

For marketing and agency operations, this translates directly. Brand guidelines, approved copy, compliance rules, regulatory requirements and product catalogues should all live in structured knowledge stores that agents query selectively. They should not sit inside system prompts that agents have to read in full on every call. 

3. Asset Libraries as Ground Truth

When agents have access to a library of pre-approved assets, they do not need to generate everything from scratch. This could include approved headlines, compliant disclaimers, cleared imagery descriptions, product descriptions, brand-approved terminology, campaign messaging, legal statements, and localisation rules. 

Instead of inventing new content every time, the agent can retrieve approved elements and compose them into the right format for the task. 

An approved asset library reduces risk. The agent is no longer starting from an empty page. It is working from a known-good inventory of content that has already been reviewed, approved, and cleared for use. 

This is faster and cheaper than generation, because the agent is reusing existing material rather than creating and checking everything from scratch. But the bigger benefit for users is control. You’ll find, in this use case, that compliance is built into the process from the start, rather than added as a final review step after the content has already been created. 

4. Tool Use and Integration

The ability to call external tools is what separates an AI chatbot from an AI agent.  

A chatbot mainly responds with text. Whereas an agent can take action. It can query databases, read live data from APIs, execute code, run OCR on documents, send notifications, create records in downstream systems, or update information inside business platforms. 

A large portion of marketing and agency work happens across project management systems, approval workflows, digital asset libraries, CRM platforms, etc. Without an agent having the ability to connect to those systems, it can only advise the user. 

A balance is needed, too. As agents gain access to more tools, the system can become harder to manage. If every available tool is included in every prompt, the prompt becomes longer, more expensive, and more confusing for the AI model. 

This is where Tool RAG comes in. 

Tool RAG applies retrieval techniques to tool selection, rather than document retrieval. This is an emerging breakthrough in scalable agent design. Rather than including every available tool in every prompt (which inflates token cost and confuses the AI model), the agent retrieves only the tools relevant to the current task. Research shows this triples tool invocation accuracy while reducing prompt length by half. 

5. Self-Correction and Self-Healing Loops

Production-grade agents need a validation step before sending outputs directly to users or downstream systems. 

A lightweight validator agent checks the output against defined rules, brand compliance, format correctness, regulatory requirements, and factual consistency. Failed checks trigger a targeted re-prompt rather than a full re-generation. For example, it can check whether a campaign summary uses approved terminology, whether a disclaimer is included, whether the output follows the right structure, or whether a claim needs human review. 

More sophisticated systems implement self-healing loops where the agent detects its own failures, diagnoses what went wrong, writes that lesson to persistent memory, and applies the correction on the next iteration. Over time, this compounds: the agent gets better at the specific tasks it performs for a specific organisation, without retraining the underlying model.  

Production deployments have demonstrated 38% reductions in infrastructure spend and meaningful quality improvements through this mechanism. The cost benefit comes from avoiding unnecessary full regenerations, while the quality benefit comes from catching and correcting errors earlier in the process. 

6. Prompt Caching

Prompt caching is a way to reduce the cost of repeated AI calls by reusing context that does not change. Many agent workflows rely on the same background information every time they run. 

If that information is sent to the model on every call, the organisation pays to process the same context again and again. 

With prompt caching, instead of treating every call as completely new, the API can store repeated prompt content so it can be reused in later calls at a lower cost. This is especially useful in high-volume workflows where the same agent runs many times with consistent instructions.  

For agencies running the same agent workflow across dozens of clients, or enterprise teams running daily batch jobs with consistent instructions, prompt caching alone can fundamentally change the unit economics of AI deployment. 

Agent performance is also shaped by the orchestration layer around the model: retrieval, tools, memory, validation loops, caching, and context management.

That orchestration layer, or harness, decides what information the model receives, which tools it can use, how it checks its work, what it remembers, and what it can reuse.

So, here’s a good rule to follow: choose the harness first, then choose the model.