What an agent needs to carry out work independently

A chat window answers a question and waits for the next one. An agent does something different: it receives an assignment, converts it into steps, calls on systems along the way, and delivers a result at the end without anyone having approved every intermediate step. That difference seems small, but it completely shifts the question. With a chat window, the question is whether the answer is correct. With an agent, the question is whether the organization around it can bear what happens if the answer is not correct.

What happens under the hood

An agent that carries out work usually runs through a chain: it reads input, consults one or more sources, makes an intermediate decision, performs an action in another system, and reports the result. Every link in that chain is a place where something can get stuck. A source that has not been updated. A system that does not accept the action. An intermediate decision that falls just outside the intended scope. How those links connect to each other, and what happens if one of them breaks, is precisely what chains in which steps follow one another is about.

Along the way, the agent often makes something resembling a decision: which of two routes, which priority, which follow-up. That is no longer a chat answer but a choice with a consequence. What an organization must arrange for that — which choices an agent may make independently and which it may not — is the subject of decision support. Without that framework, an agent carries out whatever the model considers most likely at that moment, and that is not the same as what the organization intends.

The question nobody asks until it goes wrong

Who sees that an agent has done something wrong, and when? An agent that carries out work without oversight only becomes a problem the moment someone notices, and that moment is often far removed from the error itself. Monitoring is therefore not an extra layer on top of the agent, but a precondition for putting it to work on anything that matters in the first place. What it takes to recognize deviant behavior before it causes damage is described under monitoring and signaling.

Where the agent stops and who takes over

Few agents do their work alone from start to finish. At some point the result has to go to a department, a colleague, a next system. That handover moment is often the weakest point: the agent delivers something in a format or at a pace that does not fit on the other side, or the recipient does not know that something is coming. How that handover is arranged, so that work does not remain stuck between the place where the agent stops and the place where a human continues, is addressed under coordination between departments.

The basic material an agent works with also deserves attention. Before an agent can carry out an action, it often first has to understand what is written in a document, email or meeting record. What that requires in terms of structure in the sources themselves is covered under reading and summarizing documents. And when an agent has to initiate an action following a conversation, it matters whether that conversation has already been recorded somewhere in a way a system can read — see recording and following up on conversations.

Why the foundational layer comes first

An agent that carries out work independently relies on three things that must already be in place before the agent exists: an organization that knows who remains responsible for what when a task is taken over, an IT infrastructure that can let systems communicate without anyone manually retyping data, and data management that ensures the agent has something reliable to read. These are not wishes for later. They are the dimensions that precede the question of whether an agent can do anything sensible. An organization can purchase excellent agent models and still find that nothing lands, simply because the foundational layer does not yet support what the dependent layer places on it.

That is also why this piece does not explain exactly how an agent is built or which tasks it can best take over. That question only comes up once it is established that the ground beneath it holds. An organization that wants to deploy agents without first knowing whether the foundational dimensions are in order runs the risk that the agent functions but the organization does not keep up — or the other way around.

What this means for the order

Agents that carry out work are not a goal in themselves but a form of automation that requires more than a chat function. They require chains that do not break, decision space that is clearly delineated, oversight that notices deviations, handovers that do not stall, and sources that are readable for a system rather than only for a human. All these components in turn depend on the foundational layer: organization, infrastructure, data management. That order is not a matter of preference but of how it works.