Every Agent Needs a Job Description

Key takeaways
  • Capability is not permission, so decide what an agent may touch before it can open a project file.
  • A one page job description covers purpose, scope, off limits, inputs, output, reviewer, and what to do when unsure.
  • An agent given a thin brief will do competent work that is wrong for you, exactly like a freelancer given a thin brief.
  • Name one human reviewer with the authority to reject, and fix the job description whenever the output misses.

TechCrunch reports that Adobe is releasing a local MCP that lets other chatbots like ChatGPT, Claude, or Gemini use Adobe features, with apps including Photoshop, Illustrator, Premiere and After Effects. The detail I keep returning to is a quote from Adobe VP Deepa Subramaniam: "We are still figuring out where and when and how to draw the line." I take her at her word. If the vendor has not drawn the line, the creative lead has to.

Who draws the line when an AI agent can work inside your creative apps?

The creative lead does. A vendor can ship the capability, but it cannot know your brand, your clients, or what a bad day looks like on your team. An agent that can open a project file needs a scope, a brief, and a named human who reviews its output, the same as any freelancer or new hire.

I spent about seventeen years building and leading a creative team at Full Sail University, and I came up as a producer, director, editor, and brand-systems designer. The rule I tried to hold to was simple: talent did not earn access to a master file. Responsibility did. Someone had decided what a person was accountable for, shown them the standard, and agreed to look at the result.

That order matters. Access came last, not first. A new editor got a brief and a bounded piece of the work before they got the keys to anything that could break the whole project. An agent wired into your creative apps has skipped straight to the keys. That is not the agent's fault. It is a gap in how we onboard, and the gap belongs to whoever runs the team.

What goes in a job description for an AI agent?

A short written page that says what the agent is for, what it may touch, what it must never touch, and who checks its work. It is the same page you would give a new hire, cut down to what matters. Without it, the agent's job is whatever the last prompt said, and that is not a job.

Here is what I would put on the page:

  • Purpose. One sentence on the job. If you cannot write it, the agent is not ready for it.
  • Scope. The files, folders, and apps it may open, and whether it may only read or may also change things.
  • Off limits. Final masters, work in client review, and anything not named in scope.
  • Inputs. The brief, the brand standard, and the references it works from.
  • Output. What it hands back, in what form, and where it lands. A copy for review, not an overwrite of the original.
  • Reviewer. One named person who looks at the output before anyone else does.
  • Escalation. What it does when unsure. Stop and ask, not guess.

None of this is exotic. It is the onboarding every good team already carries in someone's head. Writing it down is the work.

Why does an agent need a brief the way a freelancer does?

Because ability without context produces work that looks right and is wrong for you. A freelancer with strong skills still needs the audience, the tone, the deliverable, and the limits. An agent is the same. It will do something competent with whatever it is given, so the brief decides whether competent also means correct.

I directed and edited long enough to know that most bad work is not bad craft. It is a thin brief. When a good freelancer misses, the first question I ask is what I told them. That question applies to an agent without any changes.

The brand system is the backbone of that brief. If an agent can open the project but has never been shown the standard, it will fill the gaps with guesses. Any capable worker, human or not, does the same thing when nobody has told them where the edges are. The fix is not a smarter worker. The fix is a clearer page, handed over before the first task.

Who reviews what the agent produces?

One named human, not a team inbox and not the agent itself. That person looks at the output before it moves forward and answers for it afterward. Review is not a courtesy step. It is the point where a tool's output becomes the team's work, and someone has to own that handoff by name.

On the team I built, work had an owner. When a piece went out the door, someone could be asked why it looked the way it did. "The tool did it" is not an answer a client or a colleague accepts, and it should not be one you accept either.

Three habits make the review real:

  • Give the reviewer the time to look, not a glance between meetings.
  • Give the reviewer the authority to reject, with no penalty for slowing things down.
  • When the output misses, fix the job description as well as the file, so the same miss does not return.

That last habit is the one people skip. A miss is information about the page you wrote.

I use AI in creative work, and I let judgment lead it. Judgment starts before the first prompt, with a decision about what the agent is for and who answers for it. Adobe has not drawn the line yet, by its own VP's account in TechCrunch. Write the page before you hand over the keys.

Frequently asked

What is a job description for an AI agent?

It is a short written page that says what the agent is for, which files and apps it may open, what is off limits, what inputs it works from, what it hands back, who reviews that output, and what it does when unsure. It is the same onboarding page you would give a new hire or freelancer, trimmed to one page.

Why can a creative lead not wait for the software vendor to set the limits?

A vendor can ship a capability, but it cannot know your brand, your clients, or your team. TechCrunch reports that an Adobe VP said the company is still figuring out where and when and how to draw the line. That line is about what the software can do, not what your team should let it do. Those limits are yours to set, and writing them down is the fastest way to do it.

Who should review what an AI agent produces?

One named person, not a shared inbox and not the agent itself. That person looks at the output before anyone else sees it, answers for it afterward, and has the authority to reject it. A reviewer who cannot say no is decoration. Choose someone who knows the brand standard well enough to see a miss quickly.

Tyler GarnerVP of Brand & Creative at Hillpointe. Award-winning creative leader in Orlando and Winter Park, FL, building brands, high-performing teams, and creative operations at scale.AboutLinkedInBook a talk
Back to The CutGet in touch