TL;DR — Key Takeaways
- Prompt engineering is becoming less important as AI platforms increasingly handle task decomposition and context management.
- Users are submitting shorter prompts, while AI systems are making more LLM requests behind the scenes to complete tasks.
- Agentic AI is shifting the focus from carefully crafted prompts to clearly defined intent.
An analysis of 3.8 million messages published by Elvex, a provider of a platform for building and deploying artificial intelligence (AI) agents, suggests that prompt engineering is starting to fade as a discipline as more end users embrace agentic and context engineering.
The report finds end users are now submitting shorter, simpler messages, with the median message length dropping to about 127 characters. At the same time, however, the average number of requests made to a large language model (LLM) rose from five to eight.
That shift is being driven by end users who, rather than trying to combine a series of prompts, are increasingly relying on LLMs to decompose tasks into a set of smaller subtasks that can be completed more reliably, says Sachin Kamdar, CEO of Elvex.
Simultaneously, platforms are making data more accessible to AI agents in a way that provides more context, he adds.
Finally, the underlying AI platform now captures the outcomes as a set of reusable AI skills that make it possible to share with other end users, thereby reducing the number of prompts and messages that would otherwise need to be sent, notes Kamdar.
As that transition occurs, the level of expertise required to derive value from an AI investment is also decreasing. Instead of requiring deep prompt engineering skills, a larger percentage of end users are now knowledge workers who are more focused on intent rather than specific prompts, adds Kamdar. In fact, end user satisfaction with AI interactions has increased overall from 60% to nearly 79% as a result, the report finds. ”End users are letting the platform do the work,” says Kamdar.
The issue, of course, is that AI agents are not mind readers, so end users still need to pay careful attention to how they frame their intent, adds Kamdar. Otherwise, the AI agent is likely to consume more tokens than necessary. Additionally, AI agents will aggressively explore any and all means to complete a task, so end users need to be as explicit as possible to prevent unintended consequences.
Naturally, a significant percentage of end users continue to refine their prompt engineering skills, but that percentage is starting to decline. Even the most advanced users of AI are starting to appreciate the reasoning capabilities that have been baked into the latest generation of LLMs. There might still be instances where a unique prompt makes it possible to customize a workflow, but over time, those instances are becoming fewer and farther between.
None of this means AI is replacing humans so much as it shines a spotlight on how the relationship between AI and individuals is evolving. Arguably, the first consideration most knowledge workers now start with is how AI can be used to complete a task, with humans then filling in any gaps in reasoning that might emerge. The challenge and the opportunity then becomes how best to harness AI tools and platforms in a way that generates a set of reusable outputs that organizations can reliably trust to perform a task multiple times within a specific set of confines.