The first AI agent I ever built was also the first one I deleted. I spent more than two months trying to train her into something useful before I finally accepted that she was not going to work properly. I have now built eight agents in total. Six have been deleted, and two have survived and are working, which puts my current success rate at 25 percent. I am continuing to build more, but I have not deployed any of those newer builds yet.
Deleting her changed the way I build agents. I am now willing to kill a weak build early when the corrections are not moving it toward the job, because I have already learned what it costs to spend months pounding my head against the wall trying to force the wrong build to work.
That is not how I see digital employees fitting into an organization. A digital employee should have a job, limited authority, access to the right tools and information, work that can be measured, and a human being responsible for its performance. If it cannot earn its place in the organization, it should not remain there merely because somebody spent two months building it.
What is a digital employee?
A digital employee is an AI agent assigned to carry a specific area of work inside an organization. It can receive instructions, use connected tools, draw from persistent organizational context, act on a schedule or trigger, and produce work that another person or system depends upon.
That makes it different from the chat window most people first encounter when they begin using AI. A chatbot waits for someone to ask a question. A digital employee is expected to recognize when work needs to begin, follow the operating instructions for its role, complete the task, and report what happened.
What makes a Hermes agent different from a basic automation?
A basic automation follows a predetermined path. A Hermes agent can use language, memory, tools, and reasoning to complete work that contains some variation. The point is not to make the agent behave like a person. The point is to give it enough room to carry work that would be too loose for a simple chain of fixed commands.
Anthropic draws essentially the same line between workflows, where models and tools follow predefined code paths, and agents, where the model directs more of its own process and tool use. Its guidance is refreshingly practical: begin with the simplest system that can do the work, then add agentic complexity only when the result is worth the additional cost and risk.
An automation might move an attachment from an email into a folder. A Hermes agent can be trained to determine what arrived, place it according to the organization’s rules, flag uncertainty, and report the action. That flexibility is useful, but it also creates risk. The more room an agent has to interpret, the more carefully the organization must define what it knows, what it may touch, and when it must stop and ask for a human decision.
Why do digital employees need job descriptions?
Digital employees need job descriptions because an agent cannot perform a role the organization itself has not defined. A vague instruction such as “help with marketing” does not establish responsibility, authority, priorities, output standards, or a stopping point.
The organization needs to know what work belongs to the agent, what information it requires, which tools it may use, how success will be measured, and which decisions remain human. Without that structure, the agent will fill gaps with inference. Sometimes it will confidently drive the work into a ditch.
Most businesses already have at least one task that gets routed through the same employee every time something strange happens. The written process covers the normal path. That person carries the exceptions in their head. Giving the task to an agent before pulling that judgment into the light does not remove the dependency. It hides it behind a new interface.
I do not believe the first question should be, “What can this agent do?” The better question is, “What job are we prepared to trust it with?”
Why have I deleted six AI agents?
I have deleted six AI agents, but the first one changed the way I approached every build that followed. More than two months of training had failed to make her work properly. Because she was my first build, I did not yet know when continued training was responsible and when I was simply attached to the time I had already spent. Deleting her finally broke my attachment to the idea that enough effort could rescue every bad build.
That deletion opened the door for me to terminate later builds much earlier. I am now quicker to get rid of an agent when it is not showing that it is moving in the right direction. I would rather lose the beginning of a build than spend another two months trying to force something that is not meant to work into becoming useful.
That does not mean I delete an agent the first time it makes a mistake. Training is part of the work. The distinction I watch is whether the corrections are making the agent more reliable and moving it closer to the job. If the same problems keep returning and the role is not taking shape, persistence begins to look less like discipline and more like attachment.
The two agents still working are not trophies, and the six I deleted were not wasted work. A dead agent can still leave behind a clearer map of the job by exposing a role that was too broad, incomplete instructions, disorganized information, or work that required judgment that should never have been delegated.
How much autonomy should a digital employee receive?
A digital employee should earn autonomy through repeated, observable performance. It should begin with a narrow assignment, limited access, clear escalation rules, and enough human review to expose where its instructions break under real conditions.
I do not believe autonomy should be granted because the first demonstration looked impressive. Demonstrations usually occur on clean inputs and the builder’s preferred path. Businesses operate through exceptions, contradictory requests, missing information, and customers who refuse to follow the script. The agent should prove it can carry the ordinary work before it receives authority over the unusual work.
The National Institute of Standards and Technology makes the same point in the drier language of risk management: results observed in a controlled environment may not match the risks that emerge in real-world use. Its framework calls for defined accountability, clear responsibility, measurement, and continued management throughout the life of the system.
Are digital employees supposed to replace human employees?
I do not see digital employees as clean replacements for human employees. I see them carrying bounded, repeatable responsibilities that consume human time while still requiring consistency, context, and follow-through.
Small businesses have people spending expensive hours checking inboxes, moving information between systems, assembling reports, and performing the same keyboard work every week because the business has never had another pair of hands. That is where a digital employee may create leverage.
The standard should not be how many people the organization can remove. The standard should be whether the agent returns more human attention than it consumes. If it requires constant rescue or pushes errors onto the people already carrying the business, it has not reduced the workload. It has changed the shape of it.
How should a small business implement its first digital employee?
A small business should give its first digital employee a narrow role, a measurable outcome, and a controlled path to greater responsibility. Start with work the organization understands well enough to teach and important enough to matter, but not so dangerous that one mistake can injure a customer, employee, or the business itself.
The first implementation should establish:
- the job the agent owns;
- the trigger or schedule that starts the work;
- the organizational information it may use;
- the tools and accounts it may access;
- the expected output;
- the conditions that require human review;
- the record used to evaluate performance.
This is less exciting than telling an agent to run the company. It is also much closer to how dependable work gets built.
Why am I becoming a certified agent builder?
I am preparing to train with Peter Swain as a certified agent builder because AI Actualized is moving toward the implementation of digital employees, particularly Hermes agents. I want the technical discipline required to build them well, but I am just as interested in the judgment required to decide where they belong.
I am not entering this training under the impression that an agent can be dropped into a company and left to raise itself. The build is only the beginning. The harder work is fitting the agent into the organization without giving it vague authority, dirty information, contradictory instructions, or access it has not earned.
My own businesses have been patient zero for this work. That lets me absorb the failed builds, broken instructions, awkward outputs, and occasional acts of digital stupidity before I place that risk inside somebody else’s operation. I have no interest in discovering what not to do at a client’s expense.
What will AI Actualized provide through digital employee implementation?
AI Actualized is preparing to help small businesses define, build, train, and integrate digital employees around work that already needs to be done. That means beginning with the role and the conditions surrounding it, not with an agent looking for something impressive to do.
For Sacramento-area businesses, the local relationship also matters. Implementation requires conversations about how work moves, where information lives, which exceptions break the process, and what the owner is prepared to delegate. Those details rarely survive a generic intake form without someone noticing the gaps.
The objective is not to install AI everywhere. It is to identify where a digital employee can carry useful weight, prepare the organization to support the role, and put enough control around the work to keep it aligned.
What makes a digital employee worth keeping?
A digital employee is worth keeping when it performs a necessary job reliably, operates within its authority, makes its work visible, and gives back more human time and attention than it takes to manage.
My current success rate is 25 percent: eight agents built, six deleted, and two still working. I expect that number to change as I continue building and eventually deploy the agents now in development. What I do not want is an artificial survival rate created by refusing to terminate anything I build.
If digital employees are going to work beside us, they should have to earn the job.
That leaves the owner with a less exciting question than what AI can do: Which job in this business is defined well enough to teach, measure, and eventually trust to someone else?
Works Cited
- Anthropic. “Building Effective Agents.” Anthropic Engineering, 19 Dec. 2024.
- National Institute of Standards and Technology. “Artificial Intelligence Risk Management Framework (AI RMF 1.0).” NIST AI 100-1, U.S. Department of Commerce, Jan. 2023.
