Test and Playground covers the ways to try an agent before you ship it — the per-agent Test panel that runs your working configuration or a published version, and the separate Playground surfaces for experimenting with tools and models.
Overview
Use Test on an agent to open a chat panel and talk to it directly. You send a message and the agent responds live, just as a user would experience it — so you can sanity-check its answers, tone, and tool use before deploying. While you're building a Simple agent, the same live Test Agent chat sits right beside the form so you can test as you edit.
Draft or Published — Which Configuration Runs?
Both, depending on what you pick. By default a test run is flagged as test mode and exercises your current working configuration — including edits you haven't published — which is what makes the test/refine loop work. If you select a specific published version to test, the run instead loads that frozen snapshot exactly as it was published, so you can verify the exact thing your users get. The Simple-agent side chat goes one step further: it can try the values currently in the form, so you can check a tweak before even saving it.
Test conversations are real runs — they create a session and traces like any user chat — so you can open the trace immediately afterwards in Observability.
What You See While Testing
Alongside the agent's reply, the test view shows its trace — the reasoning and tool calls it made to produce the answer. This is what makes testing useful for debugging: you can see which tools the agent invoked, what knowledge it retrieved, and where an answer came from, then adjust the instructions, tools, or knowledge accordingly. Test, read the trace, refine, and test again until the agent behaves the way you want.
The AI Agent Playground
The AI Agent Playground is a separate scratchpad for experimenting with tools and actions. It has two panes: on the left you pick an app, action, and connection to assemble a tool; on the right is a live chat to try it. The Playground talks to a built-in demo agent rather than one of your own, so it's a safe place to explore how a tool behaves and how the agent calls it before you wire that tool into a real agent. Your conversation is preserved as you switch between the tool setup and the chat.
Test Panel vs Playground
Use the Test panel to validate your agent — its instructions, knowledge, tools, and a specific version — before deploying. Use the Playground to experiment with a tool or action in isolation against a demo agent, without touching a real one. In short: Test is for the agent you're shipping; the Playground is for trying things out.
The Playground on the AI Models Screen
The AI Models screen has its own Playground, opened with the Compare Models button in its header. It has two tabs — Compare and Chat — and it's a different tool from both the per-agent Test panel and the AI Agent Playground above: those try your agents and tools, while this Playground tries the models in your workspace catalog directly, letting you chat with them or line up their specs and pricing side by side before you commit one to production. Switching between the Compare and Chat tabs keeps whatever models you've already picked loaded — you don't lose your selection moving between them. Only chat-capable models (the same models enabled for answer generation, as described in Model Library) show up in its model pickers.
Chat Tab
The Chat tab lines up live chat panels for up to four models side by side, one column per model. From an empty board, pick a model and select Start chatting; add more columns anytime with Add model, up to the four-model limit. Each column's header gives you:
A model picker to swap that column to a different model.
A sync toggle — a synced column shows a Synced pill, and typing or sending in one synced column mirrors into every other synced column, so you can prompt several models at once with one message; switch it off to type into that column alone.
A Model parameters popup with that model's inference controls — Temperature, Top P, Top K, Max Output Tokens, Frequency Penalty, Presence Penalty (or whatever settings that particular model exposes). Changing these only affects the current playground session unless you select Save model configuration, which writes them back as that model's default settings. Revert model configuration discards an unsaved change instead.
Clear conversation to reset just that column, and a remove action to drop it from the board entirely.
A Clear all conversations button resets every column at once. An agent/team picker at the top (with All, Teams, and Agents filter chips to narrow the list) lets you route the whole conversation through a chosen agent or team instead of the default built-in assistant — the same choice applies across every column. Every message you send is a real, live call to that column's model, metered the same way as any other model usage in the workspace.
Compare Tab
The Compare tab needs at least two models. Starting fresh, you pick Model A and Model B and select Compare models; if you arrive here from the Chat tab with one model already on the board, that model fills slot A for you. Once a comparison is running, add up to two more models with Add to compare, for four side by side at most — at the limit you can only swap a column's model, not add another, until you remove one; you can never drop below two.
The comparison is a table with one column per model and three sections: Overview (provider, input modalities, output modalities), Pricing (input price and output price per million tokens, plus a cache read price row when a model prices it), and Features — the same seven capability flags shown on a model's own details page (function calling, structured output, streaming, extended thinking, context caching, batch, fine-tunable). Each column header lets you change or remove that model directly. Switch on Highlight best to mark whichever column has the lowest input price and whichever has the lowest output price.
Note: If the models' details fail to load, the table shows an error with a Try again button rather than silently rendering blank data.