I Gave an AI Agent a Desk
Chatbots offer answers. Agents ask for tools, permissions, and a place inside the operation. Working with one changed the conversation from prompting software to managing it.
For the past few years, talking to artificial intelligence has mostly felt like consulting an unusually well-read stranger.
You ask a question. It produces an answer. You request a piece of writing, some code, or a plan, and a few seconds later it appears in the chat window. The results can be impressive, but the basic arrangement remains familiar: you ask, it advises, and you perform the work.
Recently, I changed that arrangement.
After upgrading some of my own infrastructure, I installed Hermes Agent, an open-source AI agent built by Nous Research. Its GitHub repository has attracted 227,000 stars at the time of writing, which gives some sense of the appetite for software that does more than talk.
I assigned it a folder on my machine where it could read instructions, create files, and leave completed work for me to review. It was a small decision with a strangely managerial feeling. I had not merely opened another chatbot. I had given the machine a desk.
An answer is not an action
The distinction between a chatbot and an agent is often made unnecessarily complicated. In practical terms, a language model can tell you what to do. An agent can be given tools and permission to do parts of it.
Ask a chatbot how many images are inside a project folder and it may explain how to count them. Give an agent access to the filesystem and it can inspect the folder, count the files, distinguish JPEGs from PNGs, identify unusually large images, and return a report.
The language model still does the reasoning, but reasoning is now part of a larger loop: understand the objective, choose a tool, perform an action, inspect the result, and decide what comes next.
That loop changes the nature of the conversation. A chatbot waits for the next question. An agent can pursue an objective across several steps. It can search, read, compare, create, test, revise, and report back. Depending on its tools, it may work with files, browsers, terminal commands, databases, APIs, schedules, or other software.
This does not make it an employee, despite the language often used to sell these products. But it does begin to resemble delegated work. And delegated work introduces a question that ordinary chat rarely requires: how much authority should this thing have?
The first prompt is a job description
When I started working with Hermes, I found that the initial prompt mattered differently. In an ordinary chat, a prompt is usually a request. With an agent, it begins to resemble a job description. You are not only describing the desired result. You are establishing the working environment, the available tools, and the limits of the assignment.
The agent needs to know what it may inspect, where it may write, what a completed job looks like, and which actions require approval. That last part is the important one.
An agent that cannot touch anything is little more than an elaborate adviser. An agent with unrestricted access to a working system is an unnecessary act of faith. The useful territory lies between those extremes.
My preferred arrangement is simple. The agent may inspect projects, read files, run diagnostics, perform research, prepare reports, and create work inside its assigned space.
It must ask before modifying production code, restarting services, deploying websites, deleting files, installing software, sending messages, changing security rules, or accessing sensitive credentials. The exact boundaries will vary. The principle should not. Autonomy is useful. Authority should remain deliberate.
The work is larger than the prompt
The appeal of an agent becomes clearer when the assignment contains several ordinary but connected steps.
Consider a web-development project. New client assets arrive in a folder. Someone needs to organize them, inspect the existing website, determine which pages lack content, optimize the images, prepare drafts, run the site locally, check its links and forms, test its performance, and summarize what still needs human attention.
None of those steps is particularly remarkable. Together, they consume an afternoon.
A capable agent could move through much of that sequence, stopping at predetermined points for approval. It would not replace the designer's eye, the developer's judgment, or the client's knowledge of the business. It would remove some of the mechanical distance between receiving the materials and making a decision about them.
The same pattern appears in server administration: inspect the services, review storage and resource use, check for failures, read the relevant logs, identify unusual behavior, suggest repairs, perform only the safe pre-approved actions, and record what happened.
Or in sales: find businesses matching a defined profile, examine their public websites, identify specific deficiencies, separate plausible prospects from indiscriminate leads, gather public information, and prepare individualized outreach for review.
The value is not that an agent can perform one clever trick. It is that it can carry context through a chain of otherwise disconnected tasks.
Most businesses do not need a robot. They need a clerk.
The public discussion around AI tends to favor spectacular demonstrations: software writing other software, machines operating computers, autonomous companies humming along without employees. The more credible opportunities are less theatrical.
A construction company does not necessarily need an artificial executive. It might benefit from an agent that reads incoming bid documents, extracts deadlines and requirements, and identifies what is missing before an estimator begins pricing the job.
A dance studio might use one to track recital music, costumes, performers, and scheduling conflicts. A tax practice could use one to maintain client document checklists and identify which returns are waiting for a W-2, a signature, or an answer. A millwork shop might use one to turn drawings into structured job sheets covering dimensions, materials, finishes, hardware, and unresolved questions.
These are not attempts to replace an industry's central expertise. They support it. The agent takes responsibility for the administrative tissue surrounding human judgment: the small acts of collection, comparison, organization, and follow-up that quietly occupy much of a working day.
Delegation reveals the quality of the manager
There is an uncomfortable lesson hiding inside all of this. Giving instructions to an agent quickly exposes whether you understand your own process.
It is easy to say, "Handle the new project." It is harder to define where the project begins, what information matters, which standards apply, what the agent may change, and how it should recognize completion.
Humans routinely compensate for ambiguous management through experience, social awareness, and questions asked in the hallway. Software is less graceful. It tends to reveal every assumption that was never written down.
In that sense, an agent is not only a worker. It is a mirror held up to the operation. If I cannot explain the workflow, the workflow may not be as organized as I imagined. If every routine task requires an exception, perhaps it is not yet routine. If I am unwilling to delegate any part of it, the bottleneck may not be the technology. It may be me.
The intelligence is in choosing the action
When people imagine intelligent software, they often focus on the quality of its answers. Agents suggest another measure: the quality of the actions they choose.
Should it search the web or inspect a local file? Should it write code or first examine the existing project? Should it continue working, ask a question, propose a change, or stop because the next action exceeds its authority?
That decision-making is where an agent becomes useful, and where it can become dangerous. A fluent answer can be wrong without leaving the screen. A poorly chosen action can alter a system, expose information, contact another person, or destroy work. The ability to act therefore makes restraint more valuable, not less.
The ideal agent is not the one that always charges ahead. It is the one that understands when it no longer has the right to proceed.
A workshop, not an oracle
I am still early in the experience, and I do not yet know which of these systems will become permanent parts of my work. What feels different is the relationship.
A chatbot is something I visit. An agent occupies a place within the infrastructure. It has a workspace, instructions, tools, memory, and a defined route by which work returns to me.
That makes it less like an oracle and more like a workshop. The results depend on which tools are hanging on the wall, what materials have been provided, and whether the person directing the work knows what he wants built.
The current AI race is obsessed with producing the smartest possible machine. In practice, the more immediate advantage may belong to the person who becomes unusually good at assigning work: defining objectives, designing boundaries, inspecting results, and retaining judgment where it matters.
I gave an AI agent a desk. Now I have to learn how to manage it.
