I posted earlier today about the limitations of agentic approaches to implementing LLMs, referencing an O’Reilly article on the topic.

This sparked an interesting question - What do you think are the major impediments at this stage? - and I want to share my full response here for easier reference. (LinkedIn’s character limits on reply threads forced me to split it earlier.)

…

Agentic approaches aren’t yet ready for prime time.

>>
We’ve been told 2025 will be the Year of Agents, but at the same time there’s a growing consensus from the likes of Anthropic, Hugging Face, and other leading voices that complex workflows require more control than simply trusting an LLM to figure everything out.

(Article URL link below)
>>

Through my work in both government and private sector settings, I’ve been extensively involved in what can be broadly classified as process optimisation and transformation initiatives.

Without exception, no organisation has 100% visibility into how tasks are actually executed. Most organisations don’t formally map their processes and those that do rarely update them consistently. Consequently, no single person or team has a comprehensive, end-to-end understanding of how work is truly carried out.

This is just for the “happy path” scenarios.

The situation becomes even more chaotic for “unhappy path” outcomes—such as edge cases or everyday use cases where processes fail or break down due to gaps in systems, people, or policies.

Further, anyone who has worked in an organisation knows there’s often a disconnect between the “official” way things are supposed to happen and the reality on the ground. Workarounds, improvised solutions, and 'kludges' to address gaps in processes, technology, or staffing are all too common.

Amid this complexity, many organisations are approaching agentic AI projects as siloed initiatives led by a single team or division. The assumption is that this team will somehow overcome its lack of a holistic organisational perspective, bridge the gaps in process knowledge or documentation, and design a fully autonomous solution capable of making decisions, managing tasks, handling exceptions, and solving problems seamlessly—all without producing any unintended negative outcomes.

It’s an old adage, but it holds true: speeding up flawed processes simply leads to making poor decisions more quickly.