All notes

Give an agent a job description

Define the records it can read, the actions it can take and the point where a person should take over.

7 min

An agent is software that can take actions with tools. That sentence is the value and the risk. A window that answers questions is a different product from software that can open a ticket, write a record, or send a message. The second one needs the kind of clarity you would give a new hire.

Before we connect a tool, we write a job description. The test of the description is simple. A supervisor who has never read a prompt can tell you whether it is right.

Five lines, filled in

The title is a job, not a personality. “Intake clerk for facilities requests” is a title. A human name and a backstory is a costume, and costumes make reviews harder because people feel rude criticising them.

The reading list is the systems and documents it may see. Anything not on the list is out of bounds. This is also how a small company like ours stays out of the rest of your network.

The allowed actions are verbs. Classify. Ask one clarifying question. Draft a work order. Verbs that are not on the list are not implied by enthusiasm.

The forbidden actions are written down, especially the efficient ones. Do not close the request. Do not spend money. Do not email the tenant. A person sends. This is how the agent earns a wider permission later, after a few quiet weeks.

The unsure path is a destination. When the record is incomplete, the agent assigns the item to a named human queue and attaches a reason. “I was unsure” is not a reason. “Building code missing, two possible sites” is a reason.

Then the model

Model choice is real work, and it is second. A fluent model with an empty job description will improvise, and improvisation with tools is how incidents start. A plainer model with a sharp job will hand work back, which is what you wanted on the hard cases.

Keep a trace. Every tool call, every draft, every override by a person. The supervisor should be able to open yesterday and see why a sentence changed. An agent you cannot review will be switched off the first week someone is embarrassed by it. That is the correct outcome, and it is avoidable.

A sample you can steal

Title
Intake clerk for facilities requests
Reads
The request form, the building list, the SLA table
May
Classify the trade, ask one clarifying question, draft a work order
May not
Close a request, spend money, email a tenant
If unsure
Assign to the human queue with the missing fact named
Done
A draft work order with its sources, or a handoff with a reason

Replace the nouns with your own queue. If a line stays blank, that blank is the project. Filling it is cheaper than connecting a write API and hoping.

The useful agents sit in a chair someone can point at.

The companies that get value here are not the ones with the longest prompts. They are the ones who can point at a queue, a permission, and a person who still owns the outcome.

If this is the situation you are in

Tell us which systems you use, which step takes the most time and what data you can share for testing.

Discuss a project