Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

Build AI agents that do one job

Don’t just give your AI agent a job title, write a succinct and actionable job description.

[Photo: Getty Images]

Profile photo of Mara Dimofte

Co-founder · Jade

The best AI agent I’ve seen has a 14-word job description. The worst one comes with a four-page instructions manual.

Every consumer brand I talk to is planning to add agents. The pilots are everywhere. Most of them won’t survive their first quarter in production, and the postmortem is almost always the same. The reason they failed wasn’t the AI model. It’s how the agent was scoped.

Give the agent one job

Industry conversations about agents are stuck on the wrong axis. Bigger models, longer reasoning chains, more tools, more autonomy. No one is focused on the question that often decides whether an agent ships: Did anyone write down what it’s supposed to do, and what it’s not supposed to do, with enough precision that you could fire it for missing the bar?

In my experience with consumer brands in the $10 million to $50 million range, the agents that work best have only one job. They keep a written list of things they refuse to do. They pull from a single data source. They might have a title, “marketing copilot,” but there is an actual job description, “drafts product descriptions for new SKUs in this brand’s voice, capped at 80 words.”

Most teams never write a job description; they only write job titles. These are the agents that fail.

More scope should mean more value. But scope is exactly what keeps an agent stuck in development, and users never get to see it. Every extra job is more to build and more that can break, so it takes longer to get right, and longer to ship. And an agent that never reaches production is worth nothing, however capable it looks on paper. Narrow agents are small enough to ship, so they’re the ones actually delivering.

The agents that ship and continue working in production have three things in common.

1. Their job description fits on a Post-it

      Take a working agent and try to write its purpose in one sentence. If you can, you probably have a successful agent. If it takes a paragraph, the scope is already too big.

      The 14-word job description I mentioned does one thing: it rewrites product descriptions for SKUs tagged “new arrival” in the brand’s voice, capped at 80 words. That’s it. It doesn’t touch subject lines, performance analysis, or pricing, even though the merchandise team keeps asking for it.

      The failure mode is the inverse. The team writes a vague mandate, something like “help the merchandise team move faster,” and asks the agent to interpret it at runtime. The ambiguity compounds. The agent ends up inheriting every disagreement your team had and never resolved.

      Take every agent on your roadmap and write its job in one sentence, with input, output, and constraint. When the sentence comes out fast, it’s a sign that the agent has a chance. The ones where you stall are telling you something.

      2. They refuse out loud

        Working agents have a written list of requests they won’t fulfill. When one comes in, the agent declines instead of guessing. It says the request is out of scope, or that it doesn’t have the data, and routes the conversation to a person. Broken agents try their best on every prompt, which sounds polite but is the failure mode.

        I watched a brand roll out a service agent that resolved 60% of tickets cleanly while quietly inventing shipping dates on the rest. The first 60% showed up in the weekly dashboard. The other 40% showed up two weeks later as chargebacks, refund requests, and a thread in the company’s internal escalation channel on Slack.

        Agent refusal is a trust mechanism, not a weakness to engineer around. An agent that says “I don’t have that data, routing you to a human” earns more trust in a single exchange than an agent that answers everything in twenty.

        Write the refusal list, then red team it. Hand it to someone who likes breaking things and watch what they pull out of the agent.

        3. They read from one source of truth

          If your agent has to reconcile Shopify, your email platform, your help desk, and a spreadsheet, it will pick the wrong source at the worst moment. Agents that work pick one store as canonical and treat everything else as derived.

          This sounds like an engineering detail, and it isn’t. It’s where most agent projects quietly die. Reconciling across systems is hard for humans and worse for language models because they will pick whichever version of the data lets them write a confident sentence.

          Sit down with whoever is sponsoring each agent and ask which system is the source of truth. If they hesitate, the agent doesn’t have one yet. The work in front of you isn’t building the agent. It’s deciding which system is the source of truth.

          Constraint leads to success

          Scoping AI agents is hard work. It requires teams to make unpopular choices.

          To founders, the narrow scope feels like thinking small. They want the grand version. Selling them on the 14-word agent instead of the four-page one isn’t a product conversation, it’s a people one. That conversation usually takes longer than the build.

          Refusal lists demo badly. They feel like admitting weakness in a room full of investors, and they look small next to the agent that confidently answers everything. But you’re trading a better demo for an agent that doesn’t invent shipping dates. Make that trade.

          The single source of truth requires an upstream data cleanup that nobody wants to do. Deciding which system is canonical means telling some department their tool isn’t authoritative anymore. It’s a political fight, not a technical one. Teams that have the fight end up with agents that work. The rest stay in pilot mode.

          Final thoughts

          Most of the energy in AI right now is going into capability. Bigger models, longer context, more tool use. That work matters, but it isn’t where the production gap is. The teams shipping working agents aren’t the ones with the best model. They’re the ones who decided early what their agent wouldn’t do, refused to expand its scope under pressure, and made the boring decision about which database gets to be canonical.

          Model quality is plateauing. Tooling is commodifying. The differentiator now is judgement about scope—and that’s as unglamorous as it sounds. The teams actually in production all have it.

          Mara Dimofte is co-founder of Jade.