Skip to main content
building your agent in relay

​
What you will learn

Designing an agent on paper is one thing. Getting it running is another. This section walks you through the actual build process in Relay, the platform your cohort has access to, so you understand what it looks like to go from a job description to a working, testable agent.

Relay fundamentals

Understand how agents, skills, and flowcharts work in Relay

Build from a prompt

Turn a natural language description into a working agent workflow

Test and iterate

Run your agent, review the output, and give feedback that improves it

Human-in-the-loop

Know when and how to inject human review points into an automated workflow

​
How Relay works

Relay is built around a simple model: you create an agent, give it a job title and description, connect the tools it needs, and then teach it skills. Each skill is a specific responsibility, one task the agent performs on your behalf. What makes Relay different from prompting a chatbot is how those skills are structured. Instead of writing a long prompt and hoping for the best, Relay turns your instructions into a visual flowchart. You can see every step the agent will take before it runs.
This matters for trust. One of the biggest concerns leaders have about agents is predictability. You do not want an AI sending emails or filing documents on your behalf without understanding exactly what it is doing. The flowchart gives you that visibility.

​
The core concepts

Agents

Your AI team members. Each agent has a job title, a description, and a set of connected tools (Gmail, Google Drive, LinkedIn, Slack, and others). Think of this as the role definition.

Skills

The responsibilities within that role. A personal assistant agent might have skills for email triage, meeting briefings, and RSVP reminders. Each skill is a separate workflow that the agent performs.

Flowcharts

The visual representation of each skill. They show the trigger, the steps, the decision points, and the output, all laid out so you can verify the logic before the agent runs.

​
Building an agent step by step

Here is what the build process looks like in practice, based on the competitive researcher example from your live session.
1

Create the agent

Give it a job title (Competitive Researcher) and a one-sentence job description (Create regular competitive analysis reports). Select the tools it needs, in this case LinkedIn for research and Gmail for delivering the report.
2

Add a skill

Based on the tools you connected and the job description you wrote, Relay suggests skills the agent can perform. Pick one, like a weekly competitor scan, or write your own.
3

Write the prompt

Describe what you want in natural language. For example: “Every Monday, research top competitors by looking at their company LinkedIn posts and the posts of their CEO. Then email me a competitive analysis report summarising their latest product updates, launches, and customer momentum.”
4

Answer clarification questions

Relay will ask you follow-up questions to fill in gaps: who are the competitors, what are their LinkedIn URLs, what time should this run? Give it the most specific information you can.
5

Review the flowchart

Relay generates a visual flowchart showing every step. Check that the trigger is correct, the data sources are right, the analysis step makes sense, and the output goes where you want it.
Relay builds the flowchart from your natural language description. You interact with it through chat: describing what you want, answering questions, and giving feedback. The flowchart is there for you to verify, not to manually construct.

​
The 10/90 rule

Getting the initial flowchart built is roughly 10 percent of the work. The remaining 90 percent is testing, reviewing, and refining. This is the same dynamic you would have with a new team member. Their first week on the job, you would not expect perfect output. You would review their work, give feedback, adjust your brief, and let them improve over time.
Agents work the same way. The first output is a starting point, not a finished product.

​
How to test and give feedback

Once your flowchart is built, hit Test Workflow. The agent runs through every step and produces its first output. Review the output the way you would review work from a direct report. Is the structure right? Is it too detailed or not detailed enough? Did it miss something? Then give feedback in plain language. You do not need to edit the flowchart manually. Just tell the agent what to change: 
Got the first report. Keep it focused on just two competitors with one
section each. I really care about product feature details, indications
of who their target customers are, and any interesting go-to-market
motions. Format it nicely with emoji for readability.
The agent translates your feedback into changes to the flowchart and the underlying prompts. Then you test again.
Be specific about what you liked, what was missing, and what you want changed. “Make it better” is vague. “Add direct quotes from the LinkedIn posts with links to each one” is actionable. The more precise your feedback, the faster the agent improves.

​
Human-in-the-loop review

Not every action should be fully automated from day one. Relay lets you inject human review points into a workflow, places where the agent pauses and waits for your approval before continuing. This is particularly valuable for actions that have external consequences: sending emails on your behalf, publishing content, filing documents, or updating records.
In the personal assistant example from the live session, the agent drafts RSVP reminder emails for meeting guests who have not responded, but it does not send them automatically. It pauses and shows you the draft. You approve or edit, and then it sends.
As you build trust in a specific workflow and verify it is producing reliable results, you can remove the human review step and let it run fully autonomously. Start with the guardrails. Remove them when you are confident.

​
Beyond the first skill

Once your first skill is running, you can add more to the same agent. A competitive researcher might start with a weekly LinkedIn scan and later add a monthly pricing page comparison, a quarterly messaging analysis, or a trigger that fires whenever a competitor’s website changes. You can also create entirely new agents for different areas of your work. The Relay team shared that their own staff manage personal assistants, LinkedIn marketers, business analysts, recruiters, sales assistants, and more. Each team member manages between 4 and 20 agents depending on their role.
You do not need to get there overnight. Start with one agent, one skill, one task that saves you time. Build from there.

​
Quick checkpoint

You are done with this section when you can:

Navigate Relay

Create an agent, connect tools, and add a skill using the chat interface

Read a flowchart

Look at a Relay workflow and understand the trigger, steps, and output

Test and iterate

Run a test, review the output, and give feedback that produces a measurably better result

Set guardrails

Know where to place human-in-the-loop review points and when to remove them

Next: This Week's Challenge

Put it all together and build your first working agent