On Using AI better: Using Workflow Automations vs Agents

The adoption of Large Language Models (LLMs) has surged since the release of ChatGPT. But in practical applications, LLMs are still rarely able to handle the entire job-to-be-done — addressing and resolving a customer ticket, finding and booking a reasonably-priced airline ticket, or managing inventory based on sales confirmations.

Previous SlideNext Slide

Automation workflows

Many have turned to workflow automation to handle multi-step processes. Workflows are directed graphs that chain actions and decision points and are used to automate more complex tasks. One of the poster childs for this is a workflow automation tool called n8n.

n8n is a low-code solution that allows for impressive multi-step automation flows that can be built in a short period of time. It enjoys a massive adoption across AI consultancies, and developers. For instance, companies use n8n to enrich their leads sheets by automating a process to scrape the socials of their leads, summarising it with an LLM, and adding the enriched data to the database.

Such automation workflows excel when their environmental factors can be well controlled or anticipated. Their counterparts are assembly lines – known input shape, known output shape.

Say you want to automate your LinkedIn posts with daily LLM slop. The corresponding workflow might look like this:

If you prioritize consistency over quality, that’s a fairly suitable use case: The Google and OpenAI APIs are reliable. Inputs and outputs can be well predicted and the workflow after some tuning – like removing those suspicious emojis 💡🔍🧠✅🔧📊🤖🎯🎉 – is likely to produce the desired result.

Where workflows break

The issue with such automation pipelines is that they become brittle as complexity grows. Conditions and exceptions are handled as new branches of the workflow logic. For simple use cases this is not a problem, but as more edge cases need to be considered, so grows the number of branches.

The same applies across individuals and teams: more workflows mean more maintenance to keep them up-to-date, to avoid logical redundancies across them, and documentation overhead.

In low‑complexity environments, operation‑focused workflows can make for quick wins. But in higher-complexity scenarios, benefits of workflow operations can easily plateau.

Agentic reasoning: Tool-centric, not pipeline-centric

An emerging approach to lift the complexity ceiling for automation is to mimic human reasoning capabilities.

What does that mean? Imagine you’re given a new task today, increase sales of your team by 200%. How would you approach it?

First, you would ensure that you a) understand both the request, as well as b) any potential references that are part of the instruction. In this case, it might be helpful to clarify things like “Does sales refer to gross margin, number of new customers acquired, or total contract value?” and “Are we comparing against last year’s performance or last week’s performance?”.

Since 200% is an ambitious target, you might ask further questions to understand the source of this expected growth should come from. “Are we launching new products or services that are expected to drive this growth? Are we expanding into new markets?”

Assuming that’s now clear, you would then make a plan on how to reach the objective. For example you might make a plan that breaks down the year into quarters; the first quarter cross-selling to existing clients, the second quarter building out referral incentives, the third quarter and so on.

In complex settings such planning is commonly explicit, and done iteratively on multiple levels (from high-level to low-level plans) or iteratively - like in agile project management. The goal is aimed at being achieved through interdisciplinary teams, requiring the combination of diverse skills.

Now you – or rather the team – would begin executing step by step as planned. But reality is often messy; the world market could see a sudden tariff bombshell announcement, referral incentives might be less effective than you imagined and so on. In some scenarios, continuing as planned isn’t an option. Depending on the obstacle, different parameters – like larger referral incentives –, even a new approach, consulting with supervisors or stopping entirely might be necessary.

This fairly intuitive approach is exactly what agentic reasoning imitates.

The Agentic Workflow as the Orchestrator

In an agentic model, workflows are repurposed. Instead of being a rigid, step-by-step checklist, a workflow is now a high-level orchestrator for a smart AI agent. The "agentic workflow" gives the AI agent freedom to dynamically figure out how to complete tasks on its own, but provides a safe and structured framework to guide its process. The workflow orchestrates the AI's thinking by guiding the overall plan, enforcing safety limits, and tracking progress.

In short, the workflow orchestrates the agent's reasoning process. It provides the structure that separates the 'what to do' (planning and reflection) from the 'how to do it' (tool execution), ensuring the entire process is safe and reliable.

In this model, tools are the core unit of capability and the orchestrator can only be as effective as the tools it has access to. Just as a project manager assembles a team of experts – a designer, an engineer, a data analyst – the agent’s reasoning loop orchestrates tools to achieve a goal. Each tool has a clear contract to perform a specific function, like retrieve_document or send_email. The agent, acting as the project manager, doesn't need to know how each tool works, only what it does. By decomposing a larger problem into smaller subtasks, the agentic model can use its versatile 'toolbelt' of composable, reusable capabilities to iteratively problem solve.

Conclusion

Choosing the right automation strategy requires balancing immediate needs with long-term goals. While workflow automation can quickly automate predictable processes or validate automation potentials, their inherent rigidity creates a ceiling on the complexity they can handle.

Agent-oriented design offers a path forward by shifting the focus from rigid, process-centric pipelines to a flexible, tool-centric model. By treating tools as reliable, composable capabilities and the workflow as the orchestrator for an agent's reasoning loop, companies can automate a new class of complex, dynamic work.

Below is a high-level comparison of the two approaches and their trade-offs: