HomeCoursesPodsCertifications
Open app
Target Audience
Skills
Submission Type

About

Suggested Prerequisites

Rubric

CategoryWeightWhat we're evaluatingNeeds ImprovementMeets ExpectationsExcellent

Certifications and courses designed for marketers in the AI era

Learn

  • Home
  • Courses
  • Pods
  • Certifications

Resources

  • Profound Partners
  • Documentation
  • Blog
  • Verify Diploma
  • What's new

Company

  • Careers
  • Profound

© 2026 Profound

    Certifications
    Beginner

    Agent Engineering

    Prove you can build, test and run a production-ready agent with AI Marketer. Give it a clear job and goal, real brand context, deliberate permissions and a trigger that runs it on its own.

    Submit for certification
    Agent Engineering
    Target Audience

    Marketers who build and run agents in Profound

    Skills

    Agent Engineering, Context & Permissions, Automation

    Submission Type

    Project Based Assignment

    About

    You’ll give an agent one real marketing job, build it with AI Marketer, and set it to run on its own. Then you’ll submit the agent and a recorded walkthrough of it in action.

    The Introduction to Agentic Marketing course teaches each step:

    • Scope a job and brief it like a new hire
    • Add brand context
    • Connect only the tools you need
    • Set approvals and access
    • Improve the agent one change at a time
    • Start it on a schedule or when something happens

    This certification asks you to apply those steps to a problem you choose and explain each decision.

    We don’t test whether you watched the training. We test whether your agent does a job well enough to hand over.

    Requirements

    Your agent must meet all six requirements. Each one names something we should see in your walkthrough.

    1. A scoped job with a business goal. One job for a named reader. Say what decision that reader makes with the output. Pick work you’d otherwise ask for again and again. Include at least one step where the agent uses judgment, such as selecting, ranking or drafting from data, not only forwarding it. State the goal as an outcome, not a task. “Write a brief” is a task. “A brief that gets us cited when buyers ask AI about this topic” is an outcome. Define a success check someone could apply to the output.
    2. Instructions that run without you. Give the agent a role, its inputs and numbered steps in order. Give every optional input a default, either on the input or in the instructions. AI Marketer writes a first version; show it in version history and name at least two edits you made. A template used as it was installed counts as unedited.
    3. Context that changes the output. Connect at least one source of context: a knowledge base (for example, a crawl of your site), your Brand Kit, a skill, or a Profound document or sheet. Tell the agent how and why to use each one. Compare outputs from the same input and name one effect you can check, such as a topic it skipped or a term it used.
    4. The right data and tools, and nothing extra. Use the Profound data that suits the job. Turn off the capabilities the job doesn’t need in Advanced settings. Connect a tool only if the job must act outside Profound; a doc or sheet delivered inside Profound needs none. Check the Allow, Ask or Block setting of every action the agent uses and every action with side effects. Say what would go wrong with a different setting. These permissions live in your tool settings and apply to every agent that uses the same connection.
    5. Tested and improved. Test with a real input before you publish. Take your first output from the earliest run in run history and name its defects. Make at least two changes, each fixing a defect you named, and fix the biggest problem first. Rerun the same input to show each defect is gone.
    6. Ready to run on its own. Publish the agent and add a trigger that fits the job. Choose a schedule, a monitor, a Profound data trigger or an event from a connected app. In Advanced settings, under Behavior, turn off Allow questions and set When a tool needs approval to Approve or Deny. Approve runs every action that would have asked, without review, so allow only the actions the job needs and block the rest. Approvals you click while testing don’t carry into triggered runs, so show a run the trigger started that finished on its own. Triggers store fixed input values, so have the instructions work out any input that changes between runs, such as a date. Deliver the output where the reader uses it, in a size they can use. Share the agent with the people who need it, or keep it private, and say why.

    How We Evaluate

    We evaluate judgment, not feature coverage. Each criterion is scored from 0 to 3 against a checklist.

    • A sprawling agent with a vague goal, every capability left on and no changes after testing won’t pass.
    • A focused agent with a clear goal, deliberate permissions and visible changes after testing will.

    What You Submit

    Two items.

    1. Your exported agent file (JSON). Open the agent in Profound, open the more actions menu (the three dots) in the agent builder and select Export. Upload the .json file it downloads. If you don’t see Export, ask your workspace admin for access.
    2. A 5 to 10 minute walkthrough video. Show the agent in action and explain the decisions behind it. Some settings live outside the exported file, so show them on screen. Your walkthrough must show:
    • The problem, the reader, the decision they make with the output, and how you check success
    • AI Marketer’s first version in version history, and the edits you made
    • Your first output, from the earliest run in run history, next to your final output from the same input
    • The defect each change fixed
    • Your context sources and one effect you can check in the output
    • Your tools and capabilities, and the Allow, Ask or Block setting for every action the run uses and every action with side effects
    • Your agent’s Behavior settings
    • Your published agent’s trigger, and a run it started, opened from run history
    • Who the agent is shared with, and why

    Settings you describe but don’t show don’t count as evidence. You can cut or speed up long waits while the agent runs. Label the cut, and keep the run’s start and its finished output on screen. Explain the reasoning behind your agent; don’t tour the interface.

    What Passing Looks Like

    Passing walkthroughs explain the purpose of the agent. Failing ones describe what the agent does.

    • A failing walkthrough: “I asked AI Marketer to build a weekly report agent, and it wrote the setup. Here it runs. Here’s the report. I connected Gmail so it sends.”
    • A passing walkthrough: “My first draft said ‘summarize what competitors published.’ I rewrote it for our sales lead, who decides each Monday which competitor claim to answer in calls. The first output listed every post, so I told the agent to rank posts by how often buyers ask AI about that topic and keep the top three. I added our Brand Kit so the answers use our product names, and the rerun now says ‘Workspace’ instead of ‘platform.’ It runs weekly with no one present, so Allow questions is off. It only creates a doc and emails the link, so those two actions are set to Allow and every other Gmail action is Block. Approvals are set to Deny, so anything unexpected fails instead of running. The sales lead reads on a phone, so the email is a link, not the text.”

    The second walkthrough shows what we score:

    • A reader and the decision they make
    • A defect in the first output and the change that fixed it
    • An effect of the context you can check
    • A permission defended by what could go wrong
    Submit for certification

    Suggested Prerequisites

      Course

      Introduction to Agentic Marketing

    Rubric

    Submissions are evaluated across 5 dimensions, each scored on a 0-3 scale
    CategoryWeightWhat we're evaluatingNeeds ImprovementMeets ExpectationsExcellent
    Goal & Instructions25%Can you give an agent one scoped, repeatable job and instructions it follows without you?The goal is a task rather than an outcome, the agent does several jobs, or the instructions are AI Marketer’s first version or an installed template with no edits.The agent does one job for a named reader, with a success check and instructions that give a role, ordered steps and defaults. One thing is missing or unexplained: the judgment step, or the named edits to AI Marketer’s first version.The agent does one job for a named reader, and the walkthrough says what decision that reader makes with the output. At least one step uses judgment, such as selecting, ranking or drafting from data. The instructions include a success check someone could apply to the output. They give a role, ordered steps, and a default or fallback for every optional input. Version history shows AI Marketer’s first version, and the candidate names at least two edits they made to it.
    Context20%Can you give an agent context that changes its output in a way you can check?A context source is connected, but the instructions never mention it, or the outputs show no effect from it.A context source is connected, the instructions say what it’s for, and outputs from the same input show a difference. The candidate doesn’t name an effect anyone can check, or the only effect is cosmetic.At least one context source is connected: a knowledge base, the Brand Kit, a skill, or a Profound document or sheet. The instructions say what each source is for. Outputs from the same input show the context’s effect, and the candidate names one effect anyone can check, such as a topic skipped or a term used.
    Tools & Permissions20%Can you connect only what the job needs and defend every permission?Tools or capabilities are left on by default, every action is set to Allow with no reason given, or the permissions are never shown.Only the tools the job needs are connected, unused capabilities are off, and the walkthrough shows the permissions for the actions the run uses. The riskiest action isn’t named, or the reason for its setting is missing.Only the tools and Profound data the job uses are connected. An agent that delivers a doc or sheet inside Profound needs no connector. Capabilities the job doesn’t use are off in Advanced settings. The walkthrough shows the Allow, Ask or Block setting for every action the run uses and every action with side effects. The candidate names the riskiest action and says what a different setting would cause.
    Testing & Iteration20%Can you critique an agent’s first output and improve it with deliberate changes?Only one change, only trivial edits such as a typo or a rename, a different input for the second run, or no first output shown.The first output comes from the earliest run in run history, and version history shows at least two changes beyond a typo or a rename. The candidate doesn’t tie each change to a defect, doesn’t explain the order of the changes, or doesn’t rerun the same input.The first output comes from the earliest run in run history, and the candidate names its defects. Version history shows at least two changes beyond a typo or a rename, each fixing a named defect, and the first change fixes the largest one. The same input is run again, and each named defect is gone.
    Unattended Operation & Sharing15%Can you make an agent run with no one present, deliver output its reader can use, and control who can use it?There’s no trigger, the agent isn’t published, or the setup would stall: questions are on, approvals are set to Ask, or the agent relies on approvals granted while testing.The agent is published with a trigger that fits the job. Allow questions is off, and approvals are set to Approve or Deny with a reason. At least two of these are also shown: a finished run the trigger started, inputs that stay correct between runs, output sized for its reader, and sharing with a reason.The agent is published with a trigger that fits the job. A run that trigger started is opened from run history, and it finished with its output shown. Allow questions is off, and When a tool needs approval is set to Approve or Deny, consistent with the tool permissions and explained. Any input that changes between runs, such as a date, stays correct. The output lands where the reader uses it, in a size they can use. The walkthrough shows who the agent is shared with and why; keeping it private is fine when the reason is given.