Picking your agentic AI toolset: a guide for the perplexed

One of the joys of navigating around a new industry buzzword is that nobody can ever agree on exactly what it means. “AI agents” are a case in point. 

Fundamentally, I think it makes sense to imagine agents sitting along a spectrum of “agency” defined by how much autonomy they exercise over the task at hand. At one end you find the humble “agent-on-rails”: a deterministic workflow with the occasional LLM call to handle an ambiguous branch. At the other end you find the “high autonomy agent”: a vaguely tasked entity, or indeed a crew of such entities, left to plan, execute and verify a process with little human handholding. Somewhere in the middle sit “specialised agents”: task-focused helpers like code assistants, built for narrow jobs but allowed freedom in how they get them done. 

This spectrum is useful. It gives us a way of talking about agentic frameworks not as magic oracles but as engineering choices. You don’t pick “AI agents” as a category—you pick an agent with the right degree of autonomy for the task you care about. In this essay I’ll use three real-world agentic AI platforms - n8n, Lindy and CrewAI - as working examples. They neatly represent different points on the spectrum: the agent-on-rails building framework (n8n), the specialised (in this case, “assistant-like”) agentic framework (Lindy), and the high autonomy, multi-agent system toolkit (CrewAI). 

To make things practical, I’ll frame the discussion around two questions: when to use them, and when not to.

The Agent-on-Rails Framework (n8n) 

Suppose you have a well-understood business workflow. It’s been running for years. You know the edge cases, the exceptions, the routing rules, the required approvals. In this case, why on earth would you hand responsibility to a fully autonomous agent? It would be madness. Unless your workflow is trivial, it will not plan and execute this process as well as you do. 

Here the agent-on-rails approach makes most sense. You painstakingly map the process: being careful to delineate the deterministic steps with well-defined activities from the fuzzier or judgement-based steps (where you’ll eventually add a sprinkling of LLM magic). This is the essence of n8n — a low/no-code platform where you drag nodes, connect APIs, and express the logic explicitly. Every path is deterministic, auditable, and visible. 

n8n started life as a pure workflow automation tool. The advent of LLMs has given it super-powers because now, you can feasibly automate workflows that contain elements of complexity well solved by LLMs but arduous to code deterministically, by hand. 

When to use it: when the workflow is well defined, structured, and predictable. Think: 

• Linking up your SaaS tools, pushing data and cases between them; 

• Compliance-driven workflows like on-boarding; 

• Customer support Q&A; routing & triaging tickets. 

Basically: a good choice for automating all the back-office plumbing that quietly keeps companies alive. 

When not to use it: when the task is messy, conversational, or requires judgment beyond Boolean rules. n8n can call LLMs, but it will never negotiate a meeting with a client or write a nuanced report. If you try to force it into these roles, you’ll spend your time building absurdly complex flows that collapse under their own weight. 

Try it out: n8n.io

The Specialised Agent (Lindy)

Now imagine the life of a sales rep or an operations manager. Their work is a patchwork of emails, scheduling, quick decisions, CRM updates, small bits of context-gathering. The workflows aren’t codified in a process manual - they live in muscle memory. Here a graph won’t help you. What you want is a digital assistant: something that behaves like a human colleague who can take a messy request and just handle it. 

This is Lindy’s domain. It is not a process automation engine but an AI assistant that plugs into your inbox, calendar, Slack, CRM. You describe what you want (“triage my inbox and draft replies”, “schedule meetings with these clients”, “keep Salesforce up to date”) and it acts with a degree of autonomy. Not infinite autonomy: it won’t redesign your strategy or invent a new go-to-market plan. But enough autonomy to save you the tedium of repetitive, unstructured work. 

To be clear: Lindy is an example of a specialised agent designed to act as a “virtual assistant”. Others exist for other tasks - for example Cognosys for data analysis or Humata AI for document Q&A. 

Each of these sits in the “middle” of the agency spectrum: narrow scope; relatively high freedom within that scope. They don’t orchestrate multi-step workflows (like CrewAI), nor are they designed for deterministic process automation (like n8n). Instead, they are engineered to be very good at one class of human task. 

When to use it: when tasks are unstructured, human-facing and frequent. Inbox triage. Meeting scheduling. Drafting routine customer replies. Logging calls and updating CRMs. Any workflow where the output is communication rather than raw data, and where real-time responsiveness matters. 

When not to use it: when you need strict process enforcement or auditable, compliance-heavy flows. Lindy is quick and flexible, but it is not deterministic. You don’t use it to enforce a four-eye check in a financial approval process. Equally, don’t expect it to orchestrate large back-office data pipelines; it isn’t built for heavy ETL jobs.

Try it out: lindy.ai

The High Autonomy Agent (CrewAI)

Finally, we arrive at the far end of the spectrum. Here you have a task that is neither codified nor repetitive, but complex and multi-step. Possibly a task which, if executed by humans, would involve a number of specialists - each performing some kind of knowledge work. “Mini-projects”, if you like. 

Enter CrewAI, a framework for deploying multiple AI agents in parallel, each with a role (sometimes called a “swarm”, in the industry). This ensemble (“crew”) can produce an output greater than the sum of its parts. 

CrewAI is definitely not a no-code, business-user-friendly framework, but it is powerful. The most impressive, agentic case studies in industry involve solutions built (in part, at least) with CrewAI. 

When to use it: when you face tasks that require reasoning, role separation, or exploration. R&D workflows, deep research agents, multi-stage problem solving. Anywhere you’d naturally assemble a human team and divide responsibilities, you can experiment with CrewAI. 

When not to use it: when the task is straightforward or repetitive. You don’t need a team of agents to sync Salesforce with your data warehouse. Nor do you want them drafting one-sentence replies to customer emails. CrewAI is powerful, but it’s overkill unless the task is genuinely complex. And be warned: it requires engineering skill. Unlike Lindy, you won’t be up and running in five minutes.

Try it out: crew.ai

In Conclusion 

What matters here is not the brand names but the framing. “Agentic frameworks” are simply ways of matching task to tool along a spectrum of autonomy. At one end, deterministic process graphs with the occasional LLM branch to save you a lot of hard-coding of logic; in the middle, specialised assistants for unstructured but routine work; at the far end, high-autonomy teams of agents for exploratory or complex challenges. 

The trick is knowing where your task sits and choosing accordingly. Too little autonomy and you drown in manual process mapping; too much and you drown in chaos. The art is to match the degree of agency to the nature of the work. 

Have you got a process in mind ready for automation? We’re experts in delivering business value using AI tools, and our Demonstrate service is designed to get results fast for your automation workflows.