What agents unlock
You just practiced telling chat tasks from agent tasks. The deeper shift agents unlock is in how you communicate with them once they are acting on your behalf: instead of translating every step of a business task into code or software commands, you can increasingly describe what you want to accomplish in natural language and let an agent carry out much of the technical work.
From giving computers instructions to giving them objectives
For most of the history of computing, humans have had to tell computers how to perform a task. A person first decides on the procedure and then translates that procedure into instructions the computer can execute: formulas in a spreadsheet, commands in a software package, or code written in a programming language such as Python or R.
Agents change this relationship. Because an agent can reason about a goal, use tools, observe the results, and decide what to do next, the human can increasingly focus on describing what should be accomplished rather than specifying every technical step required to accomplish it.
A small example: summing the numbers 1 through 9.
Traditional: write the procedure in Python
total = 0
for i in range(1, 10):
total += i
print(total)Agentic: state the objective to Claude Code
Add up the numbers from 1 to 9.The Python version tells the computer exactly how to compute the answer, step by step. The Claude Code command only states what result is wanted; the agent decides how to produce it — in this case, most likely by writing and running the equivalent code itself.
“Vibe coding”
In February 2025, AI researcher Andrej Karpathy gave a memorable name to one version of this new way of working: vibe coding. He opened the post that coined the term like this:
There's a new kind of coding I call “vibe coding”, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.
Instead of personally writing and debugging every line, he described telling the AI what he wanted, running what it produced, showing it errors, and continuing the conversation until the software worked.
The phrase captured a larger change in human–computer interaction: increasingly, you can describe an intent in natural language and allow AI to translate that intent into executable code. The human interacts with the behavior of the system while the agent handles much of the underlying implementation.
The broader opportunity
This shift matters well beyond software engineering. Many business tasks combine judgment with technical execution: analyzing data, building dashboards, preparing reports, researching markets, reconciling documents, monitoring operations, or creating internal tools. Agents can lower the technical barrier to performing this work, and allow more people to build and automate business solutions without first becoming expert programmers. The central skill shifts from telling a computer exactly what to do at every step toward clearly defining the problem, supplying the right context, designing an effective workflow, and evaluating whether the result is good enough to use.
Example: Forecasting product sales
Imagine that an intern is asked to analyze historical sales for a retailer's products and produce forecasts for the next several months.
The traditional workflow
Traditionally, the intern would have to perform the technical workflow themselves. They might begin by opening the data and figuring out what each row and column represents. They would then identify missing values or data errors, clean and transform the data, choose an appropriate forecasting method, calculate forecasts, evaluate the results, and create figures to communicate what they found.
To make the computer perform these steps, the intern would also need to know how to express them in a form the computer understands. They might write Python or R code, construct formulas and charts in Excel, or learn the commands of a specialized forecasting package. In other words, the intern not only needs to understand the business problem; they also need to translate the solution into detailed technical instructions.
The agentic workflow
With an AI agent, the intern can begin at a higher level — but a good first prompt still states what patterns to look for and what the output should look like, not just the end goal:
Goal: Analyze historical sales by product and forecast demand for
the next six months.
Look specifically for:
- Seasonal patterns, such as holiday or back-to-school spikes
- Periods where a promotion or a stockout may be distorting the
underlying trend — flag these rather than forecasting through them
- Products trending up or down independent of season
Forecast each product line separately rather than the total. Flag
any product where the forecast is especially uncertain. Summarize
the findings as one chart per product category plus a short
paragraph of takeaways I can put directly into a slide.The agent inspects the dataset, writes and runs the cleaning and forecasting code, and returns a first pass — forecasts for every product, with a note that it treated a large sales spike in late November as part of the seasonal pattern. The intern recognizes a problem and follows up:
The November spike is Black Friday, a one-time promotion, not a
seasonal pattern. Re-run the forecast treating it as a promotional
effect instead. Also exclude the three weeks in March when the
warehouse was closed for renovation — that was a stockout, not a
drop in demand. Show me the forecast before and after these changes
so I can see how much they matter.The agent revises the analysis and returns updated forecasts along with the comparison the intern asked for. Rather than manually programming each operation, the intern supervises a computer that can determine and execute many of the steps needed to reach the objective — while still catching and correcting the assumptions that matter most to the business.
The human role shifts rather than disappears
The important change is not that the human no longer matters. It is that the human's role moves up a level. Instead of spending most of their effort translating each step into computer code, the human can focus more attention on defining the objective, specifying what a good result looks like, providing relevant business context, identifying important constraints, and checking whether the final output makes sense.
In the forecasting example, the intern still needs to make consequential decisions. What time horizon matters to the business? Should the forecast optimize average accuracy or avoid large errors on high-revenue products? As the example above shows, was an unusual sales pattern caused by a promotion or a stockout rather than real demand? Which assumptions should be preserved in the forecast? An agent can perform much of the technical work, but these questions require business context, judgment, and oversight.
Exercise: What would you delegate?
Think of a repetitive analytical or operational task from a class, internship, student organization, or job. Write down:
- the business objective you are trying to achieve;
- the information and tools an agent would need access to;
- which steps you would allow the agent to determine on its own; and
- which decisions or outputs should still require human review.
The goal is not to design the technical implementation yet. Instead, begin practicing the shift from thinking about commands to thinking about objectives, context, constraints, and evaluation.