While there is a lot of buzz around AI Agents, there is also a realization that there is confusion about the definition of an agent. Anthropic came out with a paper that breaks up Agentic systems into 2 parts - Workflows and Agents.
Personally I like this definition because it creates space for both architectures and a range of possibilities in the middle.
I asked Perplexity to come up with a definition of AI Agents based on what all the big AI vendors are saying and it came up with this -
AI agents are autonomous software entities that perceive their environment, make decisions, and take actions to achieve specific goals or perform tasks on behalf of users or organizations. They can operate across multiple systems, handle complex processes, and often aim to replicate or augment human-like capabilities in specific domains
If we stick with the purists' definition of Agents as visualized above, I believe we are a ways away from this reality - at least when it comes to solving business problems.
Let me elaborate with a specific example. Let's say I want to build an Autonomous AI Agent (not workflow) to do Account Research.
Goal - Conduct account research on prospects acting as a Sales Rep at company Y. Success is that you generate an Account Plan for each prospect.
The key for this agent to be successful is access to data and this is where the challenges arise when solving business problems.
Not all data (about prospects or your company) is publicly available.
Not all companies (prospects) have the same publicly available data (some have 10Ks and some are private).
A good account plan, needs me to understand the prospects initiatives (where do we get this from).
A good account plan needs to cover the needs of different titles.
On the topic of Titles, what is the definitive source of information about Titles? C Level titles you can find on the company page but what if you need to go one level lower. What if the company has many divisions and many C level people?
Where do I get definitive information on the divisions / departments within my prospect?
The bigger question here is - What is a GOOD account plan? Is there 1 universal account plan?
There was a study done that showed that giving LLMs conflicting information can cause challenges. Guess what organizations have - tons of conflicting information about the same topic. Multiple drafts of organization priorities, multiple versions by different people of organization priorities, priorities from old executives etc. So if I feed this to an LLM and expect it to be autonomous, it will not work.
But data is not the only reason Autonomous AI Agents in Business are going to take time. Business is an Infinite Game and players keep changing, rules keep changing and there is no defined endpoint - All of this will make it a struggle for an autonomous agent that needs clear goals, clear directions etc.
I am a big fan of Simon Sinek and he talks about the Finite vs Infinite game -
I think the best analogy of autonomous agents playing something like akin to an infinite game is self driving cars. But even here, self-driving cars operate more like repeated finite games than a truly infinite game. While there is ongoing complexity, changing traffic conditions, and evolving regulations, each driving task (getting from point A to point B) still has a clear objective and (usually) well-defined rules. With all the data that has been collected, it has been made into a finite game.
For those of us who have worked in business, we know how much we as humans struggle with the infinite game - Objectives keep changing, Rules change, Stakeholders change and success is vaguely defined. Even with companies that have V2MOMs and OKRs, we have all struggled with knowing if what we are doing is having an impact, seen our projects get shelved, new priorities getting added, new leaders have their own ideas etc.
Imagine an autonomous AI agent operating in this chaos?
I will conclude by saying that there is a huge opportunity for Workflow based agentic systems. For those on the fence, this is the best we have today. Also, I see Workflow based agentic systems akin to all those humans driving cars that collected all the data that helped map streets and define rules.
Organizations that start by building workflow based agentic systems will clean up their data, get better at defining the flows that can all then be used in the future to train an autonomous agents.
I don't see it as Workflows vs Agents, I see it as Workflows are the stepping stones to Agents (if that is even realistic) and the only way to know is to start building workflows and appreciate the complexities.
I am thankful for tools like Agent.ai and AnyQuest.ai that help me build these workflow agents as it gives me an appreciation of the struggles my autonomous agent clone will face.