Is Prompt Engineering Still Relevant? [2026]

Is prompt engineering still relevant when the conversation has shifted to agents, automations, and autonomous workflows? According to AISA research, 24.7% of the 93 people starting the AI fluency assessment chose prompt engineering as the AI skill they most want to develop next, making it the second most popular answer. But it was not first. That spot belongs to building AI agents or automations, selected by 39.8% of respondents. The gap tells a clear story about where attention is moving, and what it means for anyone deciding where to invest their learning time.

1. Prompt engineering24.7%23
2. Building AI agents or automations39.8%37
3. Evaluating AI output critically17.2%16
4. Understanding AI architecture12.9%12
5. Knowing when NOT to use AI5.4%5

93 people answered, 21 September to 26 September 2026. Source: AISA research, aisa.to

Is Prompt Engineering Still Relevant in 2026?

Prompt engineering is still relevant, but it is no longer the skill most people prioritise when they have to pick one thing to learn next. Nearly one in four people (24.7%) named it as their top development goal, which puts it well ahead of understanding AI architecture (12.9%) and knowing when not to use AI (5.4%).

Prompting Is Not Dead, It Is Assumed

The question "is prompt engineering dead" keeps surfacing in forums and LinkedIn threads. The data does not support that framing. A quarter of respondents still actively want to get better at it. What has changed is that prompting is increasingly treated as a baseline competence rather than a differentiator. The people who already feel confident in prompt chaining and context engineering are looking for the next layer.

The Shift From Skill to Infrastructure

Prompt engineering started as a standalone discipline: craft the right input, get a better output. In 2026, prompts are components inside larger systems. A well-structured prompt still matters inside an agentic workflow, but the person building that workflow needs additional skills (orchestration, error handling, evaluation) that go beyond writing a single instruction. This is why the question "should I learn prompt engineering" increasingly gets the answer "yes, and."

Which AI Skill Do People Most Want to Learn Next?

Building AI agents or automations is the AI skill most people want to develop next, chosen by 39.8% of the 93 people AISA asked. When AISA posed the question "What's the AI skill you most want to develop next?", the full ranking was: agents and automations first (39.8%), prompt engineering second (24.7%), evaluating AI output critically third (17.2%), understanding AI architecture fourth (12.9%), and knowing when not to use AI last (5.4%).

Agents Dominate the Wish List

The 15.1 percentage-point gap between agents (39.8%) and prompting (24.7%) is the largest gap between any two adjacent options. This is not a marginal preference. It reflects a concrete shift in what people believe will be most valuable to them.

The Long Tail Still Matters

Evaluating AI output critically drew 17.2%, and understanding AI architecture drew 12.9%. These are not glamorous skills, but they are load-bearing. Anyone building agents without the ability to evaluate what those agents produce is building on sand. The AI skills gap statistics across AISA's broader data reinforce that evaluation and architecture understanding tend to be areas where people overestimate their own ability.

Why Do So Many People Want to Build AI Agents?

Agents represent the next step in practical AI leverage: moving from single-turn interactions to multi-step, semi-autonomous workflows that complete real tasks. The 39.8% figure reflects a desire to go from using AI as a tool to deploying AI as a collaborator that can act on its own within defined boundaries.

From Chat to Workflow

The appeal is straightforward. A prompt gets you one answer. An agent gets you a process. When someone asks "prompt engineering vs AI agents," they are usually asking whether to keep optimising individual interactions or start building systems that chain multiple interactions together. The data suggests most people are ready for the latter.

The Gap Between Wanting and Doing

Wanting to build agents and being ready to build them are different things. Agent development requires understanding context engineering, error recovery, tool use, and evaluation, skills that sit across multiple competency areas. The Stack Overflow Developer Survey 2025 found that 66% of developers cite AI solutions that are almost right, but not quite as a significant challenge. That "almost right" problem compounds in agentic systems where one step's output feeds the next. If you have experienced this frustration firsthand, you are not alone, why is AI so frustrating is one of the most common questions people bring to AISA.

What Agent-Building Actually Requires

Here is a practical breakdown of what each skill area prepares you for:

Every row feeds into agent-building. The skill people most want (agents) depends on the skills ranked below it.

Is Prompting Still the Foundation for Agents?

Yes. Every agent runs on prompts. The system prompt defines the agent's role, constraints, and output format. Step-level prompts determine what happens at each stage of a workflow. The quality of those prompts directly determines whether the agent produces useful results or cascading errors.

Prompt Engineering Inside Agentic Systems

In a single-turn interaction, a mediocre prompt produces a mediocre answer and you move on. In an agentic system, a mediocre prompt at step two corrupts steps three through ten. This is why prompt chaining (the practice of structuring sequences of prompts where each output feeds the next) becomes a critical skill for anyone building agents. The stakes of each individual prompt go up, not down, as systems get more complex.

The "Should I Learn Prompt Engineering" Decision Tree

If you want to build agents: you need strong prompting skills first. If you want to evaluate AI output: you need to understand what good prompting looks like to know when the model was set up to fail versus when it failed on its own. If you want to understand AI architecture: prompting gives you the practical context to make architectural knowledge actionable. Prompting is not the destination for most people anymore, but it remains the on-ramp.

What About Evaluating AI Output Critically?

Evaluating AI output critically was chosen by 17.2% of respondents, third place, but arguably the most underrated skill on the list. A 2025 global study by KPMG and the University of Melbourne found that 66% of respondents rely on AI output without evaluating its accuracy. That figure should concern anyone deploying AI in production, whether through prompts or agents.

Why Evaluation Gets Overlooked

Evaluation is invisible work. Nobody posts "I caught a hallucination today" on LinkedIn. But the ability to spot when a model is confidently wrong (and to design systems that catch those failures automatically) is what separates someone who uses AI from someone who uses AI reliably. AISA's own data on safety and evaluation skills shows this is consistently the area where people score lowest, as explored in AI safety skills: the lowest score.

Evaluation as an Agent Skill

For agent builders specifically, evaluation is not optional. Every agent needs validation checkpoints. Every multi-step workflow needs a way to detect when an intermediate output has gone off track. The 17.2% who chose evaluation as their top priority may be the ones best positioned to build agents that actually work in production, because they are focused on the failure modes rather than the happy path. The free critical thinking with AI guide in the AI Fluency Academy covers practical techniques for structured output evaluation.

How Do You Choose Your Next AI Skill?

Start with an honest assessment of where you are now, then pick the skill that removes your biggest bottleneck. The answer to "what AI skill to learn next" depends entirely on your current capability profile, not on what is trending.

Match Your Skill to Your Bottleneck

If your prompts produce inconsistent results, prompt engineering is your bottleneck. If your prompts work but you cannot tell when the output is wrong, evaluation is your bottleneck. If you can prompt and evaluate but you are still doing everything manually, agents and automation are your bottleneck. If you are building agents but they keep breaking, architecture understanding is your bottleneck.

The Stacking Effect

AI skills are not interchangeable, they stack. The AISA assessment measures AI fluency across five dimensions precisely because competence in one area without the others creates blind spots. Someone who can build agents but cannot evaluate their output will ship unreliable systems. Someone who can evaluate output but cannot structure a prompt will spend all their time fixing problems that better instructions would have prevented.

Get a Measured Baseline

The most common mistake is choosing what to learn based on what sounds exciting rather than what you actually need. An AI fluency assessment gives you a measured baseline across all five dimensions (prompting, critical thinking, technical understanding, workflow application, and responsible use) so you can see where your gaps are before you commit time to closing them. AISA's assessment uses evidence-linked scoring against a published rubric, and every person who completes it gets a free report showing exactly where they stand.

The 39.8% who want to build agents and the 24.7% who want to sharpen their prompting are not making wrong choices. They are making different choices based on different starting points. The question is whether they know their starting point accurately, or whether they are guessing. If you have been wondering how AI has changed your job, measuring your current skills is the first concrete step toward deciding what to learn next.

About this research

This finding comes from AISA's own research. AISA asks a short research question of people as they start the AISA AI fluency assessment, one question at a time. The question in the chart above was asked from 21 September to 26 September 2026, worded exactly as shown and with the answers in the order shown, and 93 people answered it. Each person counts once: when someone answered more than once, their latest answer is the one counted.

More findings from this research are collected in AISA research and data.

How to cite this research: AISA (2026). "Is Prompt Engineering Still Relevant? [2026]". AISA research. https://aisa.to/blog/is-prompt-engineering-still-relevant

---

Related reading: AI Skills Gap Statistics [2026 AISA Data], where the biggest skill gaps actually are. Related reading: Why Is AI So Frustrating? [2026 Survey], what people find hardest about working with AI. Related reading: AI Safety Skills: The Lowest Score [Data], why safety and evaluation consistently lag behind.

Frequently Asked Questions

Is prompt engineering still worth learning?

Yes. Prompt engineering remains the second most wanted AI skill among the 93 people AISA asked, with 24.7% choosing it as their top development priority. It is also a prerequisite for building agents and automations, since every agent runs on prompts. The skill is not dead, it has become foundational rather than standalone.

What AI skills should I learn in 2026?

The five areas people are focused on are building AI agents or automations (39.8%), prompt engineering (24.7%), evaluating AI output critically (17.2%), understanding AI architecture (12.9%), and knowing when not to use AI (5.4%). The right choice depends on your current skill profile. Start by identifying your biggest bottleneck rather than following the most popular option.

Is building AI agents harder than prompting?

Agent-building requires prompting skills plus additional competencies: orchestration, error handling, evaluation, and architecture understanding. A single prompt that goes wrong produces one bad output. A prompt that goes wrong inside an agent can cascade through every subsequent step. The complexity is higher, which is why strong prompting and evaluation skills are prerequisites, not alternatives.