Margin of Safety #3: What is an agentic process anyways?

A lot of hype around ‘agentic’, but what does it actually mean?

TLDR: Agentic systems – where LLMs control workflow – promise to automate complex tasks. True value of these agentic systems lies in tackling problems beyond traditional decision trees, and demand careful consideration of trade-offs like cost and complexity

Beware of Agentic Hype

Agentic is the buzzword of the new year. But what does it actually mean? Much of the agentic conversation is focused on outcomes – automating human processes and performing feats that have classically taken human intelligence – versus the capabilities that drive those outcomes. But without clarity on the capabilities, how can we be confident that the outcomes will materialize?

Additionally, anytime someone suggests that a single technical architecture – be it agentic, peer-to-peer, cloud, or other – is going to rule them all, we should be suspicious. Architectural choices involve tradeoffs. And it’s rare to see the choice that’s truly right for all circumstances. (Except for avoiding your least favorite programing language, because c’mon).

The current agentic hype is sidestepping the question of tradeoffs by being vague about the capabilities. But the more we’re vague about the capabilities, the easier it is for agentic to mean everything and nothing all at once. AlphaGo achieved better-than-human performance with a mix of AI techniques, including deep learning and neural networks but not including a transformer-based architecture (other engines have since experimented with transformers). It’s probably not what people are thinking of when they talk about agentic processes, yet it would qualify under many of the definitions we read on LinkedIn.

Getting Specific: Defining Agentic

So, what does agentic mean and where will that architecture be deployed? I personally like Anthropic’s proposed definition; a true agentic process is one in which an LLM is managing the control flow of the system, rather than a process in which a classically specified flow potentially invokes one or more AI methods or tools. This architecture has tradeoffs; it’ll be both higher cost and harder to test/QA than a process using a classic control flow. So – as Anthropic calls out – it only makes sense to deploy in contexts where the overall system flow is infeasible to express with classic methods. Under this definition, if someone tells you they’re writing a fully agentic system to do something that you could express as a simple decision tree, you should be very suspicious. You’re likely signing on to either an over-engineered system or to pure snake oil.

Under this definition, we gain clarity about the likely benefits of agentic processes. The gains will come when we’re automating processes that couldn’t be described with decision trees. These are often mid or senior level tasks: think about how often entry level tasks are outlined with highly prescriptive onboarding guides. The existence of those guides is a strong indicator that automation doesn’t require the level of dynamic control flow that would merit an agentic system. Instead, we should expect those tasks to be automated with simpler, cheaper techniques.

What agentic isn’t

Some sources claim that agentic processes are ones which use common LLM engineering techniques. Things like prompt chaining, LLM-as-judge and chain-of-thought prompting or reasoning all fall into this bucket. They’re examples of building blocks that allow for LLMs to be used in more complicated scenarios with higher quality results. However, they don’t have to be used in agentic scenarios – they can also be deployed in contexts where a human operator (or a simple decision tree) initiates and manages a relatively simple and singular process.

Other sources suggest that agentic processes are ones designed to replace humans. This definition is again challenging, because humans labor can be used for arbitrarily simplistic or complex tasks! Plus, we’ve already seen complex tasks automated with classical AI methods. For example, companies like Airbus and Boeing have been building increasingly complex autopilots for decades. And beyond auto-pilots, engineering is full of automated monitoring systems that have been slowly removing requirements for full time, dedicated human eyes.

The term “agentic” has become a buzzword, but its meaning is often vague and overhyped. Mislabeling simpler automation techniques, like decision trees or prompt chaining, as agentic dilutes the term’s value. Ultimately, agentic architecture should be reserved for tasks that genuinely require dynamic control, rather than straightforward processes that can be automated using simpler, more cost-effective methods.

Cutting through the Hype

You don’t have to share this definition of agentic, but if you’re talking about agentic systems, it should have some definition. And when you’re hearing about what agentic systems are going to do, it’s worth trying to understand whether the discussion is based on a shared understanding if what an agentic system is. If our experience is any guide, half of the time it won’t be.

Thoughts? What’s your agentic hot take? What do you think the key capabilities (versus outcomes) of an agentic system should be?

Stay tuned for more insights on securing agentic systems. If you’re a startup building in this space, we would love to meet you. You can reach us directly at: kshih@forgepointcap.com and jpark@forgepointcap.com.

This blog is also published on Margin of Safety, Jimmy and Kathryn’s substack, as they research the practical sides of security + AI so you don’t have to.