Field Guide · framework
What Is an AI Agent? A Field Guide Definition
An AI agent is configured intelligence that works a defined stream of work — what teammates ask for, what events fire, what runs on the calendar — and operates in your team's chat, where the work is visible to everyone.
The phrase AI agent has been used to describe so many different artifacts in the last two years that it has stopped meaning anything in particular. A chatbot with a search tool is called an agent. A workflow automation with a plug-in is called an agent. A model that has been told it is an agent is called an agent. None of these are wrong, exactly. They are just useless to a buyer who is trying to decide what to spend money on. This piece is not the canonical definition of the word, because there isn't one yet, and the part of the industry that wants there to be one will spend the next two years arguing about it. This is a working definition in the shape a buyer can actually use — what the agent takes as input, what it produces, where it operates, and what it can refuse.
Why it matters
A buyer who has not been told what an agent is, in a way that maps to the work, falls back to the marketing definition: an agent is the smart version of the AI tool we sold you last year. That definition makes every claim plausible and none of them comparable. A working definition that names the inputs, the outputs, the operating surface, and the limits gives the buyer a frame to read every Roster entry and every pitch with — what is its stream of work, where does it live, and what does it refuse. The argument-shape is more useful than the dictionary entry.
An agent has four parts a buyer should recognize.
1. A defined stream of work.
The agent does not respond to a prompt in a window. It takes as input a stream of work that hits a team — internal requests from teammates, external events from systems and the world, scheduled work that recurs on a calendar. The stream is bounded. The agent's constitution names what is in the stream and what is not. A customer service agent's stream is the support channel and the ticketing system; it is not the team's general chat. A marketing agent's stream is competitor announcements, the content calendar, and the analytics threshold alerts; it is not the founder's email.
2. A configured corpus and constitution.
The agent is not a generic model with a longer prompt. It carries a corpus that matches the role — the documents, frameworks, and reference patterns a practitioner would consult — and a written constitution that names the rules it follows when the corpus is silent and the agent has to make a call. The constitution is the part the buyer should read like a job description, because it is one. It states the agent's posture, the boundaries it will not cross, and the path it takes when it cannot proceed.
3. A working surface where the team watches.
The agent operates somewhere the team can see. For most teams this is Slack. For some it is Teams or another chat surface. For a few it is a doc canvas or an email thread. The surface matters because the surface is what makes the work auditable; an agent whose work product is private to the manager who hired it is structurally indistinguishable from a chatbot in a private window. An agent in shared sight is the part of the definition that does the load-bearing work for adoption.
4. A published list of limits and an escalation path.
Every Fidelic agent ships with a written list of what it cannot do, kept on the Roster entry, kept current, and read aloud during onboarding. The limits are the part of the definition that earns the rest its trust. A claim about what the agent can do is a marketing claim. A list of what it cannot do is a contract.
These four parts together make an agent. Strip any one of them and what you have is a different artifact — a chatbot, a workflow, a copilot, a tool. Put them together and what you have is the thing the rest of the Field Guide describes.
The edge
The reason this definition is useful — as opposed to the marketing definition or the academic one — is that it tells you what to ask. Ask what stream the agent works. Ask what the corpus is. Ask to see the constitution. Ask where the agent posts. Ask for the limit list. If the answer to any of these is unclear or absent, the artifact is not yet an agent in the sense that matters to a hiring decision. It might still be useful. It is just useful in the way a tool is useful, not in the way a teammate is. The distinction matters most when the team has to live with the choice.
Honest take
This definition is the working version, not the final version. The industry will keep moving. The word agent will keep being applied to new things. The boundary between an agent and a workflow will keep getting tested. What I am describing here is the version of the word that holds up against the question a buyer is actually asking — what am I paying for, where does it live, what can it refuse — and that is the version of the word that earns the rest of the conversation. When the industry settles on a different definition, the underlying shape will not have changed, only the name. The shape is what matters. The shape is what the four parts describe.
An agent is configured intelligence that works a defined stream, in shared sight, with a written constitution, a published list of limits, and a escalation path. Anything less is a tool. Anything more is a job description.