Do You Actually Need AI Agents? (Why I Use Folders Instead)

Summary

The video argues that many workflows labeled as AI agents can be handled more simply by giving one AI model access to organized folders of instructions and context, along with a small set of general-purpose tools.

Folder-Based Approach

For a task such as creating an investor presentation or PDF, the speaker proposes keeping relevant information in ordinary, readable files—especially Markdown or text—organized into folders. These might include:

• How you like work done, such as presentation or document preferences.
• Who you are and your company context, including brand voice and background.
• Information about the audience, such as investors or customers.
• Task-specific instructions or “skills,” such as how to make a presentation.

With a tool or “harness” that lets an AI read, write, and search files—and potentially search the web—the model can retrieve relevant material for each request. The AI effectively takes on different roles depending on the instructions and context it reads: a presentation creator, PDF creator, or marketing assistant.

Why Agent Systems Became Complicated

The older approach often involved building a separate agent for each task. A presentation agent might require:

• A database or file store containing personal and company information.
• Code and API connectors to retrieve that information.
• Tools for creating slides or interacting with presentation software.
• Web-search tools to research investors.
• Error checks to confirm the correct people or data were found.

A different output format, such as a PDF, could require another toolchain or agent. The speaker says these systems could be difficult to control and describes early AI-generated PowerPoints—based on his experience training employees at a large company—as especially poor. One workaround was to generate slides using HTML or images and then transfer them into PowerPoint. He notes that presentation tools have since improved.

The Harness and the Role of Context

The proposed alternative is a lightweight set of common tools—a harness—rather than many task-specific integrations. The essential capabilities described are reading files, writing files, and searching, including web search where needed. The model can then use those tools alongside folder-based instructions and data.

The speaker connects this approach to the way transformer models use context: the same prompt can lead to different outputs depending on what information precedes it. A folder of personal preferences, company details, and audience information supplies an additional layer of context, helping the model produce an answer that better matches the user’s intentions. The speaker emphasizes that this keeps the user in control of the context and instructions.

Historical Perspective: Language Becoming Executable

The video places the idea in the broader history of computing, describing a long progression toward turning human instructions into machine actions:

• Jacquard’s loom used punched cards to control weaving patterns, illustrating how a physical pattern could encode a sequence of operations.
• Ada Lovelace’s work on the Analytical Engine explored how a machine might manipulate symbols and potentially do more than numerical calculations. The speaker references her notes on the engine, including the famous “G note.”
• René Descartes and later Alan Turing considered what it would mean for a machine to think or communicate. Turing’s “Imitation Game” is presented as a test of whether a machine could converse convincingly, rather than a definitive test of consciousness.
• Binary and electronic computers turned logical processes into on/off signals. The speaker also mentions Turing’s codebreaking work during World War II.
• Claude Shannon is discussed in connection with information theory and the separation of meaning from the engineering representation of information.
• Grace Hopper’s compiler work is presented as a major step toward translating human-readable commands into machine operations. The speaker relates early skepticism about compilers to contemporary concerns that AI-generated code is hard to control.
• Languages such as Fortran, C, and Python progressively abstracted lower-level machine instructions. The video uses print("Hello, world") to illustrate that a simple human-readable command relies on many layers of software and hardware beneath it.

The central historical claim is that programming has repeatedly advanced by encapsulating complex lower-level processes behind simpler, more accessible instructions. In the speaker’s view, natural-language requests such as “build me a PDF for my investors” continue that progression—but the AI still needs the user’s intent, context, and preferences to know what “good” means.

Main Takeaway

Rather than creating a separate agent for every workflow, users can often create and maintain instructions and context in folders, then let one AI model use them through a small set of general tools. The speaker acknowledges that specialized scenarios may still need dedicated systems, but presents folder-based context as a simpler, flexible starting point.

Main speaker/source: Jake Van Clief.