How to Define One Job for an AI Agent That Ships

Most agents that flop share the same defect. Their job description reads like a job posting for a person: research competitors, draft the brief, format the deck, send it to the client, follow up if they go quiet. That is five jobs. An agent asked to do five jobs will do all of them badly, and the buyer will churn before you understand why.

Figuring out how to define one job for an AI agent is the single highest-leverage decision you make as a creator. Get it right and the agent runs cleanly, testing is fast, and buyers understand what they are hiring in ten seconds. Get it wrong and no amount of prompt tuning will save it. Here is the process, step by step.

Start with the finished state, not the task

Before you write a single instruction, describe what exists in the world when the agent is done. Be specific. A finished state is a file, a message, a row in a database, a decision with a reason, a scored list. It is something a buyer can point at and say "that is what I paid for."

Compare these two framings for the same rough idea:

• Task framing: "Helps with sales outreach."
• Finished-state framing: "Produces a 5 email sequence for one named prospect, personalized against their last 90 days of public activity, delivered as a Google Doc."

The second one tells you what the agent makes. It tells you when the run is over. It tells the buyer exactly what they are buying. If you cannot write the finished-state sentence, you do not yet have an agent. You have an aspiration.

Run this quick test on your idea. Fill in the blank: "When this agent finishes a run, the buyer has ______." If the blank fills with a concrete artifact and a clear stopping point, keep going. If it fills with "help" or "support" or "insights," go back and narrow.

Write the job in one sentence, then attack the ands

Now write the job as one sentence. Say what the agent takes in, what it produces, and any critical constraint. Something like: "Given a signed vendor contract, produce a redline of unusual clauses with a plain-English explanation of each risk."

Then hunt every "and" in that sentence. Each "and" is a fork. Some are fine. Some are hiding a second agent. Ask of each one: could a buyer reasonably want the first half without the second half? If yes, you have two agents.

Examples of ands that reveal a split:

• "Reviews the contract and negotiates the changes with the counterparty." Two agents. Review is analysis. Negotiation is correspondence. Buyers want them separately.
• "Builds the financial model and presents it to the board." Two agents. Modeling is a spreadsheet. Presenting is a narrative deck.
• "Screens inbound applicants and schedules the interviews and writes the rejection notes." Three agents. Different inputs, different outputs, different failure modes.

Examples of ands that are fine because they describe one continuous output:

• "Extracts the unusual clauses and explains the risk of each." One artifact, the annotated redline.
• "Reads the transcript and returns a summary with action items." One artifact, the summary.

The test is whether the ands describe one deliverable or several. If several, split the listing.

Name the one decision the agent makes

Every useful agent makes at least one judgment call inside its run. Naming that call sharpens the scope more than anything else.

For a Contract Reviewer, the call is "is this clause standard or unusual for this type of agreement." For an Equity Research Analyst, it is "which of these signals actually move the thesis." For a Market Researcher, it is "which of these competitors is a real threat versus noise." One call, applied over and over inside a single run.

If your agent seems to make three different kinds of decisions, you are likely bundling three agents. A hiring agent that decides who to interview, decides what to pay them, and decides how to onboard them is really three specialists in a trench coat. Each decision has different inputs, different reference material, and different failure modes.

Write the decision down as a question the agent answers on every run. If you can write one question, you have one agent. If you need three, you have three.

Draw the boundary out loud, in the listing

A scoped agent tells the buyer what it does and what it deliberately does not do. This sounds like a downside. It is the opposite. Buyers trust agents that are honest about their edges, because it means the agent will actually finish the job it accepted.

Put the boundary in the listing itself. A useful pattern:

1. What it does: the one-sentence job with the finished state.
2. What it takes as input: the exact files, links, or fields required.
3. What it returns: the artifact, in the format buyers get.
4. Where it stops: the adjacent work it does not do, with a nod to the agent or expert that handles it.

Example for a Vendor Contract Reviewer: "Reviews standard SaaS vendor contracts up to 40 pages and returns a redline of unusual clauses with plain-English risk notes. Does not negotiate with the counterparty, does not produce a signed final version, does not review contracts governed by non-US law. For non-US contracts, use the Multi-Jurisdiction Contract Reviewer. For negotiation support, book the human expert."

That listing costs you nothing and earns you the trust of every serious buyer who reads it. It also protects your ratings, because buyers with out-of-scope needs self-select out before they run the agent and leave a two-star review.

Test the scope against three real inputs before you publish

Scope always looks tighter in a doc than it does in the wild. Before publishing, run your draft agent against three genuinely different inputs from your actual work history. Not toy examples. Real ones with the messiness intact.

Watch for these failure patterns, because each one points to a scope problem, not a prompt problem:

• The agent asks the buyer for something the buyer would not know. The input list is wrong or the job requires expertise the buyer does not have. Narrow the input requirements.
• The output shape changes between runs. The finished state is not defined tightly enough. Pin it down: "always returns a Google Doc with these five headers, in this order."
• The agent produces something correct but off-topic. The job is broader than one sentence. Cut a branch.
• You find yourself editing the output significantly every time. The judgment call belongs to a human, not the agent. Move that part to Expert work and let the agent handle the mechanical stretch.

This is where building your first agent as a real, testable draft beats another week of planning. The failures tell you where the scope is wrong faster than any spec review will.

A concrete before and after

Here is a real-shaped example of how the collapse from many jobs to one job looks.

Before, one bloated listing: "Marketing Ops Agent. Audits your funnel, writes the campaign brief, builds the email sequence, sets up the tracking, reports on results weekly, and flags what to change." That is six agents. Buyers cannot picture the finished state. You cannot test it. Failure could come from any of six places and you will not know which.

After, six clean agents from the same expertise:

• A Funnel Audit Agent that returns a scored diagnostic of the current funnel with three prioritized fixes.
• A Campaign Brief Writer that turns a positioning doc plus a goal into a one-page brief.
• An Email Sequence Drafter that produces a 5 email sequence from a brief.
• A Tracking Setup Checker that reviews a GA4 and CRM configuration for a named campaign and returns a gap list.
• A Weekly Funnel Report Agent that pulls last week's metrics into a fixed template with commentary.
• A Campaign Change Recommender that reads a report and returns three ranked adjustments with rationale.

Now each one has a finished state, one decision, and a boundary. Buyers can hire one, two, or all six. Augie can chain them into a workflow so the report feeds the change recommender automatically, memory carrying preferences from one run to the next. Your listings get discovered because each one matches a specific search intent. And when something breaks, you know exactly which agent to fix.

That is the multiplication move. One afternoon of scoping, then six agents earning independently, each pointing to your Expert work when the job crosses into judgment.

An agent works well when its job can be stated in a single sentence with a clear finished state. If describing the job takes a paragraph with several ands, you have two agents wearing one listing. Split them, define each finished state, and publish them as separate roles that can be run alone or chained together. Ready to try it against your own workflow? Open a draft and list your agent with a one-sentence job and a boundary you would defend out loud.