Skip to main content

How Can I Create Custom Agents?

Christopher John avatar
Written by Christopher John
Updated over a week ago

Introduction

In Pencil, Custom Agents allow you to scale creative work without losing your brand voice or strategic focus. You might think of them as smart teammates - each with a specialised role, memory and context - ready to be pulled into your workflow whenever needed.

What is a Custom Agent?

A Custom Agent in Pencil is a configurable AI assistant designed to support a specific creative or strategic task. All Custom Agents are defined by:

  • Instructions: Its core behaviour and tone

  • Knowledge: Brand-specific or topic-specific context

  • Capabilities: Tools like web search or file reading

  • Model: The underlying AI model powering its responses

Agents can work independently or as part of a team or chain in a creative pipeline - depending on your use case.

How do I Build a Custom Agent?

Option 1: Duplicate a Pre-built Agent and Adapt it

This is likely to be the fastest way to get started. Choose one of the pre-built Agents and, hovering over it, click the three dots and choose 'Duplicate':

The copied agent will appear at the bottom and hovering will reveal a menu that includes 'Edit':

This is a good way to see how high-performing Agents are configured and adapt them to your needs.

Option 2: Create a New Agent from Scratch

Click 'New Agent' in the top right of the Agents homepage and you will be prompted with:

"What agent would you like to make?"

You can choose either:

  • To enter a natural-language prompt (e.g. "Help write headlines for luxury skincare ads in a playful but precise tone")

  • Skip this step and set everything manually

Prompting can jumpstart your Agent setup, but you can always refine it further in the full manual view.

If you choose to begin with a Prompt, Pencil will use it to auto-fill your Agent's name, goal and some instructions, expanding on your prompt.

To get the best outcome:

  • Be clear about the Agent's role (e.g. “concept testing for B2B SaaS landing pages”)

  • Specify the input context it will need

  • Mention tone or brand considerations

  • Clarify output format (e.g. “short headline list with variants”)

Manual Configuration (Post-Prompt or Skipped)

Once inside the setup screen, you'll configure:

Name and Description

  • Choose a clear title that reflects the Agent's purpose

  • Use the description to help your teammates understand when and why to use it

Model

  • Select from the available AI models which model will underpin your agent:

    • OpenAI GPT-4o

    • Google Gemini 2.0 Flash

Instructions

  • Define the Agent's behaviour clearly

  • Example: "Always suggest emotionally resonant headlines that follow this format: [benefit] + [emotional hook]. Reference tone given in Brand Library”

Conversation Starter (optional)

  • This helps onboard new users to the Agent, acting like a pre-filled suggestion or nudge

Knowledge & Capabilities

  • Upload files in the following formats for your Agent to reference, or write/paste plain text.

    • .pdf

    • .docx

    • .pptx

    • .txt

    • .csv

    • various image formats

  • Brand Knowledge: choose whether you want the Agent to access and reference your Brand Guidelines

  • Web Search: choose whether you want your Agent to be able to access and reference data from the web

  • URL Access: choose whether you want your Agent to be able to scrape and reference information URLs from any of the source data you have provided

Multiple Agents and Modular Workflows

Multiple Custom Agents can be used powerfully in sequence. For example:

  1. Audience Research Agent ->

  2. Messaging Strategy Agent ->

  3. Headline Generator Agent ->

  4. Scriptwriter Agent

Each Agent builds on the work of the last, bringing its own customised specialism and focus while preserving context across the creative pipeline.

Deployed into the right context, creating a Custom Agent can take minutes, and can save hours.

Start small, test the results, and refine your Agent(s) over time. The more context you give, the better the results.

Did this answer your question?