AI agents #2: Is it better to prompt or to build an agent?

March 1, 2025

Originally published in Forbes France in March 2025. As Forbes France has ceased publication, this article is rehosted here in its original form.

An article by Flavien Chervet

03/10/2025

Remember René¹: that innovation-funding consultant, a bit of a geek, who had discovered the joys of generative AI back in 2022 by using ChatGPT to help write his technical files. Then, in 2025, he had thrown himself into the great adventure of "AI agents" by automating the various stages of designing his files, from gathering information all the way to the final submission.

Yet, as he dove into the world of agents, René asked himself a fundamental question: "Why build an agent when I can just hand the AI a single, ultra-detailed super-prompt?" René is a geek. And a geek always thinks his ideas are brilliant. But luckily, a geek also experiments. So René gave the approach a try. He began drafting his super-prompt by listing all the actions to carry out: collect the documentation, list the innovation points, produce a draft file, run review passes back and forth, and so on. Then he asked his favorite generative AI to execute the whole thing in one go.

Far from being outlandish, this "super-prompt" approach caused quite a stir when LLMs (Large Language Models, like GPT) appeared, capable of handling large volumes of text. So why "waste your time" building a sophisticated agent that autonomously chains together several modular steps, when you can do it all in a single request?

The answer comes down to several points, all of which evolve fast as the technology matures. But today, in 2025, the limits of the super-prompt method remain very real:

1. Focusing on one task at a time is still often more effective

Classic AI systems (ChatGPT, Claude, o1…) still prove more reliable when they concentrate on a precise, well-bounded goal. Indeed, if you ask them to carry out a long series of interlocking tasks, with no segmentation, they risk getting lost or making consistency errors.

True, the arrival of reasoning models such as o1 or Grok3, and their strong ability to handle complex tasks, has gradually pushed this limit back (which also makes them excellent candidates to serve as the intelligence engine of agents, notably the orchestration blocks that handle planning the agent's tasks). But in practice, for many companies, and for a consultant like René, it remains safer to split the mission into successive steps, driven or orchestrated by an agent.

2. The constraints of the "context window" and the "output window"

To answer a request, an AI only has access to a certain amount of text as input (the context window) and can produce an output limited in size (the output window). Among the industry giants, these limits keep expanding: Gemini 2.0, Google DeepMind's latest model, can take up to two million tokens (~1.5 million words, or about twenty complete novels) as input. Google DeepMind even describes this model as "Built for the agentic era"². Anthropic's Claude 3.7 climbs up to 128,000 tokens (~96,000 words, or a complete novel) as output via its API.

But we are still far from an infinite prompt! Even if these figures seem gigantic, they remain insufficient to chain together the entire set of complex steps you would want to entrust to an AI, especially if each of them requires back-and-forth exchanges, the compilation of the digital reflections of many expert sub-processes, feedback loops to review, critique, improve, and so on. Hence the value of breaking the work down into different modules or micro-tasks, orchestrated by an agent.

3. Not all AIs are equal for every task

One of an agent's major advantages is being able to call on different models depending on the sub-task to be accomplished. Let's take some very concrete examples:

To write code, a specialized AI (such as Codestral³, the specialized model from the French startup Mistral, or Qwen2.5-coder⁴, the one from the Chinese giant Alibaba) will be more performant and/or less costly than a generalist AI.
To research information, an agent might prefer to rely on Deep Research (the agent recently introduced by OpenAI⁵, with an open-source version created by HuggingFace⁶) or on competitors that are experts in data extraction and synthesis.
For purely creative work, such as brainstorming or storytelling, another, more "literary" and generalist model (like Claude) will be a good fit.

In short, each model has its strengths and weaknesses, its costs and its constraints. By unifying everything under the banner of a single "super prompt," you force the use of a single model, which probably will not be the best-suited one for every step of the process. An agent, on the other hand, is able to "route" each sub-task to the right AI or the right specialized service. You gain performance and save money, so why hold back?

4. Agents also handle non-AI "tools"

This is a point many people overlook: an agent is not made only of AI blocks. It also integrates traditional tools (CRM, databases, Excel or Google Sheets files, third-party APIs like AirBnB or Stripe, and even connected objects or physical robots in a factory!).

As long as you stick to a single big prompt, you can only give the AI limited access to these external sources. It cannot navigate autonomously to, say, retrieve a script on YouTube, draw from your emails, or question another specialized AI. With an agent, the system can plan, then use all these tools to carry out its actions at the right moment.

5. Control, robustness, and reproducibility of the process

René saw it firsthand: with a super prompt, if the AI produces an unsatisfactory deliverable, you don't really know why. Was it a lack of clarity in the wording? A clumsy instruction? A logical error in the model's reasoning?

When you use an agent, you can follow each step of the reasoning chain, review the output halfway through, correct or replace the AI if there's a problem. This modularization offers finer control and a more robust process, especially for critical tasks (administrative filing, financial diagnosis…). Here you find the best of both worlds: the power of connectionist AIs (based on neural networks), with their emergent, systemic approach to knowledge and reflection, within a modular, almost symbolic structure, where each block is isolated and testable.

6. (And let's be honest…) Agents are also the trend

Yes, it's a wink, but there's a kernel of truth to it. The current enthusiasm for AI agents encourages companies to capitalize on this new approach rather than stay on the simple chatbot paradigm. In the tech industry, the race for novelty is part of the DNA, and AI agents embody the most promising frontier of the moment.

Even with the trend effect, agents have real advantages that the prompting approach lacks. It remains highly relevant to fall back on a classic prompt for simple, one-off tasks, or ones with a fairly narrow scope: quickly writing a paragraph, summarizing a short document, correcting text, translating, etc. On the other hand, as soon as it's a matter of carrying out several successive steps, involving several data sources or tools, or requiring strong consistency and continuous adjustments (iterative back-and-forth, interactions with external systems, etc.), building and using an AI agent makes full sense. The agent then offers modularity, robustness, and finer control than the single "super prompt," while making it possible to orchestrate different AIs and algorithmic actions within one and the same process.

René, for his part, keeps refining his strategic-consulting offering. His agent is gaining autonomy and is even starting to handle client follow-ups for the missing pieces of the files… no more need to write those dreadful reminder emails! Now and then, out of curiosity, he tries a super-prompt again, holding on to the hope that a next-generation model will finally manage to do it all in one shot. But for now, the agent approach remains unbeatable.

Want to go further on this topic?

Discover Hypervibe