Roadmap

Run evaluations programmatically from code with full control over test data and evaluation logic. Evaluate agents built with any framework and view results in the Agenta dashboard.

Automatically evaluate every request to your LLM application in production. Catch hallucinations and off-brand responses as they happen instead of discovering them through user complaints.

Configure LLM-as-a-Judge evaluators with custom output schemas. Use binary, multiclass, or custom JSON formats. Enable reasoning for better evaluation quality.

Define and validate structured output formats in the playground. Save structured output schemas as part of your prompt configuration.

Use Google Cloud's Vertex AI models including Gemini and partner models in the playground, Model Hub, and through Gateway endpoints.

Filter and search for traces based on their annotations. Find traces with low scores or feedback quickly using the rebuilt filtering system.

Completely redesigned evaluation results dashboard with performance plots, side-by-side comparison, improved testcases view, focused detail view, configuration visibility, and run naming.

Create folders and subfolders to organize prompts in the playground. Move prompts between folders and search within specific folders to structure prompt libraries.

Improve organization structure by adding projects. Create projects for different products and scope resources to specific projects.

Add Jinja2 template support to enable conditional logic, filters, and template blocks in prompts. The prompt type will be stored in the schema, and the SDK will handle rendering.

Add PDF support for models that support it (OpenAI, Gemini, etc.) through base64 encoding, URLs, or file IDs. Support extends to human evaluation for reviewing model responses on PDF inputs.

Create reusable prompt snippets that can be referenced across multiple prompts. Reference specific versions or always use the latest version to maintain consistency across prompt variants.

We are adding the ability to filter traces by date range in the metrics dashboard.

Analyze prompts and suggest improvements based on best practices. Identify issues, propose refined versions, and allow users to accept, modify, or reject suggestions.

Add a button in observability to open any chat span directly in the playground. Creates a stateless playground session pre-filled with the exact prompt, configuration, and inputs for immediate iteration.

We are making it easy to use and navigate in the playground with large testsets .

Using testcases from different testsets is not possible right now in the Playground. We are adding the ability to append a single testcase to a testset.

We are reworking the testset view to make it easier to visualize and edit testsets.

We are adding the ability to cache prompts in the SDK.

We are adding the ability to version testsets. This is useful for correctly comparing evaluation results.

We are adding the ability to tag traces, testsets, evaluations and prompts. This is useful for organizing and filtering your data.

We are adding the ability to use built-in LLM tools (e.g. web search) in the playground.