So much of LLM volatility and unpredictability can be solved by defining deterministic logic around LLM outputs. 

This is the premise of the enterprise AI Agent development tools report that I wrote for n8n, which explored the market of no-code/low-code automation tools with AI implementations. Agents depend on the non-deterministic outputs of large language models. These varied outputs make AI-based automation volatile, so defining deterministic logic can control the agents’ inputs and outputs to make sure it doesn’t behave in a way you don’t want it to.

The most bare-bones way of writing an AI agent is by hard coding it. If you want to use an abstracted tool to make the process quicker and easier to maintain, you can choose from the following:

1. Workflow-based builders
   1. AI native workflow builders
   2. Workflow builders with AI retrofitted
2. Non-Workflow builders

Let’s explain each of the above.

Workflow-based builders

AI-native workflow builders

These are mostly startups that have built their platform extensively (and exclusively) to build AI agents

These include the likes of Vellum, Dify, Langflow, Flowise, with their exclusive focus on building AI agents. They provide users with a great deal of control over agents’ behavior, but make it harder to integrate with their IT stack. This makes them better suited for use cases where Agents use web resources, SaaS apps, and documents rather than coordinating a bunch of on-premises enterprise applications.

Workflow builders with AI retrofitted

These vendors found themselves in a great position to hop onto the biggest bandwagon of the 2020s.

Workflow automation tools have some distinct maturity advantages, particularly around depth and breadth of integrations, which provide tried-and-tested methods of bringing a tech stack together without writing custom connectors for each application. Native Agentic AI dev tools, on the other hand, have the agility advantage and don’t have to retrofit AI to an existing product.

Writing integrations, engaging with technology partners, building a community, and offering out-of-the-box content are very long exercises. It is not an apples-for-apples comparison to look at a 2024-established Stack-AI against a 2013-established Workato in terms of the integrability features. Tools that score well in integrability and lower in codability are also great contenders for less complex agentic systems. If you don’t need to have orchestrated multi-agents working in parallel, but still need to integrate with your wider system, these tools may very well be fully satisfactory.

Non-Workflow AI-native builders

These include tools such as LangSmith, Crew AI, Restack, and Writer.com. I haven’t spent too much time evaluating these tools, so I can define it by saying how it isn’t workflow-based.

Non-workflow based agent development tools are:

• Permissive - they provide agents a greater deal of autonomy and self-determination, which is particularly useful for reasoning models
• Higher technical knowledge - they require users to have a better understanding of how LLMs work and at least a high-level knowledge of coding practices (i.e. how to define a condition, what an API key is, what’s an URL)
• Supporting novel use cases - they are best used for processes that don’t follow standard operating procedures and perhaps for ad-hoc requests. For example, asking in natural language what’s the overall customer sentiment in support chat, and comparing it with sentiment across social media.
• Mainly for internal use cases - see my original premise, I wouldn’t expose an AI agent to external use cases (Customer-facing) without some serious deterministic guardrails.

Workflow based agent development tools are:

• Restrictive - these agents still have self-determination, but they naturally give you a way of interacting with the LLM output
• Low technical knowledge - no-code interfaces are much more intuitive for citizen developers. While code-based capabilities still exist, they are mainly used as a fall-back for instances when the no-code GUI doesn’t support a use case.
• Automating pre-existing processes - writing workflows inherently requires a workflow to exist. Even a self-determining agent must be defined within a given framework or business context.
• Expose Agents to end-users - with the appropriate guardrails, a workflow-based agent can be exposed to end-users, which includes support and sales bots.

A Market Landscape Analysis

If you’re interested in learning more about how workflow-based automation tools can be used to write AI agents, you can read the Evaluation of Enterprise AI Agent Development Tools report. We assess some of the leading automation tools in the market as well as new entrants that are built with an AI-first approach.