Building AI Systems: The Case for Custom Workflows Over Agents

Over the winter holidays, Anthropic published a thought-provoking piece addressing a question that has been quietly circulating among many of us in the tech industry: are AI agents as effective as they’re hyped up to be? Or, in most cases, would a well-designed AI-powered workflow actually provide a better, more practical solution?

This debate isn’t just about semantics — it cuts to the heart of how we think about integrating AI into real-world applications. With so much buzz around AI agents, it’s worth taking a step back to ask: what exactly distinguishes an agent from a workflow, and which is more effective in practice?

Anthropic’s insights boil down to two key arguments:

1. AI-powered workflows are often more effective than AI agents.
2. Frameworks like LangGraph and Amazon Bedrock Agents can encourage over-engineered solutions — ones that are unnecessarily complex and prone to abstraction-induced headaches when debugging.

Their recommendation? Start simple. Build straightforward workflows that are transparent and heavily monitored with performance metrics. Resist the allure of building sprawling AI agent systems, especially when they obscure the underlying mechanisms and make troubleshooting difficult.

At Bartleby.dev, we agree with Anthropic’s recommendation. Our approach centers around pairing two key technologies — Large Language Models and Knowledge Graphs — to create custom workflows that are modular, optimized for iterative improvement, and built on context specific to your organization. In our experience, these workflows deliver superior results compared to generic, off-the-shelf AI agent software that doesn’t leverage organizational knowledge and isn’t built custom for your use case.

Now, let’s take a step back. How exactly does Anthropic differentiate between an AI agent and an AI-powered workflow, and why does understanding the difference matter?

Anthropic defines AI agents as systems designed to operate autonomously across a range of tasks, often with minimal human intervention. Think of them as virtual assistants that aim to “own” a process end-to-end. Popular frameworks like LangChain, LangGraph, and Amazon Bedrock Agents provide the scaffolding for developers to build these agents by connecting various tools, APIs, and models.

In contrast, an AI-powered workflow is more like a pipeline—a sequence of well-defined steps where AI is integrated into specific parts of a process to enhance its efficiency or accuracy. These workflows typically consist of multiple steps, making them easier to debug and understand while still being tailored to achieve highly specific objectives.

For example:

• An AI agent might be tasked with handling customer service queries end-to-end. The agent is responsible for creating a task plan and then looping through LLM calls to it’s conclusion, all while managing context across conversations, calling tools, escalating issues when needed, and logging outcomes autonomously.
• An AI-powered workflow might instead define explicit steps where an LLM call (or calls) assists a human by auto-summarizing tickets, suggesting responses, or flagging high-priority issues.

The distinction matters because it influences how we design and implement AI systems. When deciding which system is best for your organization, it is important to consider the trade-off between autonomy and control.

Anthropic highlights several reasons why AI agents are not the right answer for most use cases:

• Over-Engineering: Many agent frameworks encourage complex, multi-layered architectures that solve more than what the problem requires.
• Abstraction Issues: High-level frameworks that are often used in building agents abstract away the inner workings, making debugging and optimization difficult.
• Monitoring Challenges: Autonomous systems are harder to monitor effectively, especially if the metrics for success aren’t well-defined.

By comparison, workflows lend themselves to simpler, modular designs where each component is transparent and can be measured independently. When working with a new technology like AI, control and visibility are important.

As we mentioned earlier, Anthropic’s thesis aligns well with our philosophy at Bartleby.dev: start simple, iterate quickly, and measure obsessively. Here’s how we approach AI-powered workflows:

• Modular Design: Each step in a workflow is a standalone module, making it easier to test, debug, and improve.
• Custom Context: Before we build workflows, we focus on the Knowledge Graph, or digital representation of your organization, to supercharge our AI. Without custom intelligence about your organization, we don’t think any application of AI, from workflows to agents, are capable of working effectively for your business.
• Performance Metrics: From the outset, we build workflows with measurable KPIs—accuracy, speed, or cost-efficiency—so we can track progress over time.
• Iterative Improvement: Workflows are designed to evolve. Feedback loops help refine each module as new data and insights become available.

Anthropic’s article is a timely reminder to focus on practicality over hype. While AI agents have their place, especially for certain high-complexity or highly dynamic use cases, most organizations will benefit far more from custom AI-powered workflows that are transparent, purpose-built, and guided by specific organizational context in a Knowledge Graph.

At Bartleby.dev, we believe the future of AI lies in crafting workflows that deliver real-world impact without unnecessary complexity. As Anthropic suggests, building an AI-first organization isn’t about embracing the most complicated versions of new technology — it’s about using that technology in the most effective and transparent way possible.

If intelligent workflows are an area your organization is pursuing, please reach out. We’d love to work with you.