Beyond the Agent Hype

Being a product engineer who's begun to explore the 'building my own thing' path, the explosion of AI agents has been nothing short of captivating.

We have been fed with the promise, the word of honour that these supposed autonomous workers will revolutionise how we interact with technology.

But to be honest, something has always felt off to me.

I then recently came across Anthropic's article on building effective agents, and what had been nagging at me finally became clear: what if agents are just workflows in disguise? And more notably, what if that's actually a good thing?

The Agent Clickbait Complex

The current narrative around AI agents sounds almost cyberpunk-esque: "Just tell the agent what you want, and it'll handle everything!".

We've all seen demos of Claude, ChatGPT or other systems booking flights, researching competitors or even buying groceries! However, tinkering around with the tech has put me face-to-face with a crucial reality: behind many seeming autonomous agent lies a crucially orchestrated workflow.

The most effective "agents" aren't even truly autonomous! They're sophisticated workflows with the appearance of agency.

The actual magic isn't the autonomy, but rather the thoughtful choreography between the capabilities of AI and human input.

The Hidden Workflows Within 'Agents'

If you pulled up the front hood of most effective AI agent systems, you'll find:

1. Structured prompting sequences that guide the LLM through predefined steps
2. Evaluate checkpoints where the quality of the output is checked
3. Fallback mechanisms when the primary approach fails
4. Human intervention triggers at critical decision points

The above are fundamentally, a workflow, right? The agent abstraction simply hides the complexity from the user, sometimes helpfully, sometimes detrimentally.

Anthropic's engineering team acknowledges this, noting that their most effective agent implementations involve "breaking down complex tasks into smaller, manageable steps with clear sequences and dependencies."

That's workflow thinking, not agent thinking.

Why Workflow-First Thinking Matters

When we start with workflows, rather than backflipping into agents. The way we approach AI implementation fundamentally changes in various ways.

1. Clarity Over Magic

Workflows make the process clearly expressed. This transparency helps the user build accurate mental models of what the system can and cannot do, allowing you to collaborate more effectively.

2. Tailored Refinement

Workflows can be optimised to the very step. When something breaks, you know exactly where the failure occurred.

With pure agentic approaches, troubleshooting becomes a "prompt engineer and pray" exercise.

3. Sensible Autonomy

Not every step needs AI.

Workflow thinking encourages us to apply AI only where it adds value, rather than forcing everything through an agentic paradigm.

Let's use an example: a sales outreach system could be positioned as an "autonomous agent that handles all your prospecting.", but in reality, the most effective implementation likely looks more like:

1. The human defines their target criteria
2. AI generates the list of prospects, based on this
3. The human reviews and refines the list
4. AI drafts personalised outreach messages
5. Human reviews and approves the messages
6. System sends the approved messages
7. AI summarises any responses for human review

This is clearly a workflow, not a single agent, and it's better for it.

The Autonomy Spectrum

I've changed my thinking from seeing agents and workflows as different approaches, to points on a spectrum of autonomy:

From left to right: Human-driven → Augmented → Supervised → AI-driven

Most current "agent" implementations fall somewhere in the middle of this spectrum, and interestingly, the most valuable applications aren't typically on the far right.

Rather, they're actually in the middle zones! Where human judgement and AI capabilities complement each other.

Finding The Sweet Spot

As I look to explore potential product opportunities, I'm learning to ask a different question than "Could an agent do this?", rather:

1. Which parts of this process benefit from automation vs human judgement?
2. Where does the appearance of autonomy help the user, and where would it do the opposite?
3. How could a hybrid workflow better serve the actual needs?

With this shift in thinking, you open up possibilities that purely agentic approaches may miss.

Where Agents Really Shine

Something to flag, I'm not saying that agentic implementations don't have their place.

In fact, they work beautifully when:

• Tasks are well-bounded with a clear success criteria
• Failures have low consequences
• The process works better when it's humanised
• Users prefer to delegate rather than collaborate

LLM's that plan and execute complex tasks using tools like code terminals and web browsers show the beginnings of true agentic ability.

The Horizon

As someone who's exploring the founder path after years in the engineering frontlines; I always grimace at buzzwords, and AI has become a particular culprit for this.

Whether we call them agents, workflows or whatever else, the most valuable AI implementations (for now, at least) will be those that:

1. Start with a deep understanding of the human need
2. Apply AI thoughtfully, rather than gung-ho
3. Have a reasonable balance of autonomy and control
4. Can evolve based on real-world use, rather than "what-ifs"

Many of the most impactful AI products are a far cry from the all-or-nothing agent paradigm, and until research and tech evolves, I theorise it will be this way going forward for a long time.

---

This post marks the beginning of my exploration into how AI can serve real business needs, particularly for SMEs where off-the-shelf solutions often fall short. If you're curious about this in this space or have thoughts on the agent vs workflow paradigm, I'd love to hear from you!