Skip to content

Build an agent around your own task.

Create a custom AI agent deliverable with defined behavior and evaluation. Use it for a specific workflow of your own.

Start free Explore pricing

From the brief to the result.

01

Define the job

Describe what the agent should do, the context it needs and how a useful result should look.

02

Evaluate the behavior

Inspect the generated agent and its evaluation results before relying on it in a workflow.

03

Refine and reuse

Adjust the instructions and implementation through the build conversation, and keep the generated deliverable.

Custom agent builds are distinct from the built-in Workload teams, which run ongoing SEO, marketing, sales and other work.

Your knowledge. Your tools. An AI assistant you can test.

Define the job, shape the conversation and inspect how the assistant behaves.

Ground the answers

Give your assistant something reliable to say.

Build a conversational assistant around your own task, audience and source material. The AI creates its instructions, organizes the supplied knowledge and builds a chat interface. Set the tone, the questions it should answer and what it should do when the information is missing.

  • Use your brief and knowledge files to define the facts it can rely on.
  • Shape the greeting, answer style and scope of the conversation.
  • Review the generated chat experience on desktop and mobile.
Illustrative support conversation answering a returns question beside its source policy

Connect the right actions

Give it tools, and a way to hand over.

An assistant can search its knowledge and use supported HTTP tools that you define. Configure when it should offer a human hand-off, such as a request outside its scope. The platform runs the model and tool calls; the generated chat interface displays the replies and hand-off events.

  • Declare the tool inputs and endpoints the task actually needs.
  • Choose when the assistant should stop and offer a hand-off.
  • Test the surrounding workflow before relying on a hand-off notice.
Illustrative assistant offering a human hand-off with the conversation context

Evaluate the behavior

Test the questions your users will ask.

The build includes example conversations with expected answers and actions. Evaluation checks those examples alongside platform probes for unsupported answers, prompt injection, privacy, scope and tool behavior. Inspect the findings, then refine the instructions, knowledge or interface in the build conversation.

  • Review ordinary questions, unknown answers and edge cases.
  • Inspect whether tool calls and hand-off behavior match the intended task.
  • Keep the agent definition, knowledge and evaluation examples with the build.
Example behavior-review table for policy questions, unknown details and hand-off requests

The workflow

Keep the work connected.

  1. Describe the result

    Explain the audience, desired output and constraints. Refine the direction before moving forward.

  2. Review the work

    Inspect the generated files, available preview and the checks relevant to the output.

  3. Make the next version

    Ask for changes, then export or use the delivery workflow supported by that project.

Explore what’s next.

See all builders

AI Website Builder

Create landing pages, storefronts and content sites. Describe the audience and the experience, then review and refine the build.

Explore AI Website Builder

AI App Builder

Build web, mobile and desktop apps around the job they need to do. Preview the experience, refine it and prepare the appropriate package.

Explore AI App Builder

AI Game Builder

Turn a game idea into a playable project. Shape the rules, characters and world, then test and iterate on the experience.

Explore AI Game Builder
Questions & answers

A little clarity before you start.

How is this different from the built-in AI Agent Teams?+
AI Agent Builder creates a custom agent deliverable for your own task. Agent Teams are the platform’s existing workflows for work such as research, SEO, marketing, sales and messaging.
What should I tell the builder?+
Describe the agent’s task, the context it should use, its boundaries and examples of good responses. Include the tools or integrations the workflow needs and when it should hand a task back to you.
Can the agent use tools?+
Custom agents can declare tools for their workflow. The implementation and required connections must be available; evaluation includes checks that declared tools receive valid calls.
How is the agent evaluated?+
The checks cover example conversations, grounding, tone and language, declared tool calls, latency and errors. Additional probes test behavior around scope, sensitive information and attempts to override instructions.
Can I inspect why an evaluation failed?+
The evaluation record includes transcripts, grades and findings. Use that evidence to review the response or tool behavior and request a targeted change.
Can I revise the agent after the first build?+
Yes. Refine the instructions, examples or implementation in the build conversation and review the next version and its evaluation. Passing checks does not remove the need to review its behavior in your intended workflow.

What will you make?

Create a build, or put a specialist team to work on your next task.

Describe your idea Explore agent teams