Skip to main content

Platform Agnostic Integration

Spec27 is designed so new agents can be integrated through JavaScript, which makes it possible to connect many different platforms and services into the same evaluation workflow.

This is useful when teams need one product surface for agents they build themselves, agents that call internal APIs, and agents backed by third-party services.

How agent integration works​

In Spec27, an agent is stored as JavaScript that receives an input, calls the target system, and returns the output that should be evaluated.

The integration pattern is intentionally simple:

return async function process(input) {
  // call the target system
  // return the final output Spec27 should score
}

Because the contract is small, the same model can be used for many different agent backends.

What "platform agnostic" means here​

Platform agnostic does not mean every system looks the same internally. It means the integration surface is flexible enough that different systems can be wrapped and evaluated through the same product workflow.

That can include:

  • internal services exposed through an API
  • third-party model or agent providers
  • vendor tools that can be called from JavaScript
  • existing agent implementations the team has already connected

Existing integrations and new ones​

Some integrations already exist in the product, and new ones follow the same JavaScript-based model.

That means teams do not need a separate evaluation system for each platform they want to assess. They can integrate a new agent by implementing the JavaScript wrapper that calls the target system and returns the final output for scoring.

When to ask the team for help​

If you are integrating a new platform and are not sure how to structure the JavaScript agent, the quickest path is to review the existing Agents guide first and then reach out to the team for help with the integration approach.

You can also use the docs help section as a starting point for common product questions:

Example: connecting a new vendor-backed agent​

Suppose a team wants to evaluate a vendor-backed assistant alongside an internally built agent.

They can create a JavaScript agent that calls the vendor endpoint, returns the output in the expected shape, and then connect that agent to the same datasets, specifications, and eval workflows already used elsewhere in the project.

That lets the team bring a new platform into the same evaluation system instead of creating a separate testing process around it.