Playground — Chat with Models and Agents
The dashboard's /playground page lets you test a model directly via the LiteLLM gateway, or chat with a deployed agent's runtime through the API proxy.
Playground — Chat with Models and Agents
The dashboard's /playground page is the fastest way to test a model or a deployed agent without writing any code. It exposes two modes:
| Mode | What it does | Backed by |
|---|---|---|
| Model (default) | Chat with a model directly. Pick an agent for system-prompt context, optionally override the model, and chat. | POST /api/v1/playground/chat (LiteLLM gateway) |
| Agent | Chat with a deployed agent's actual runtime — the same code that runs in production. | POST /api/v1/agents/{id}/invoke (API proxy) |
Mode is persisted per browser in localStorage (agentbreeder.playground.mode), so reloading keeps you where you left off.
Model mode
Model mode is the original Playground experience. It runs the agent's system prompt + your input through the LiteLLM gateway, returning the raw model completion plus token counts, costs, latency, and any tool calls the playground was able to elicit.
Use Model mode when you want to:
- Iterate on prompts without redeploying
- Compare models for the same prompt (use the Default model dropdown to override the agent's configured
model.primary) - Save a turn as an eval case via the Save button on any assistant message
Agent mode
Agent mode chats with a deployed agent — i.e. an agent in the registry that has a non-empty endpoint_url. The dashboard:
- Lists every deployed agent in your workspace in a dropdown
- Sends your messages to
POST /api/v1/agents/{id}/invoke, which proxies to the agent's runtime - Resolves the agent's
AGENT_AUTH_TOKENserver-side from the workspace secrets backend keyed byagentbreeder/<agent>/auth-token— you never paste a token in the browser - Round-trips
session_idso the runtime can stitch turns together
No deployed agents in your workspace yet? Run agentbreeder deploy
on an agent — once the registry record gets a real endpoint_url
it will appear in the Agent mode dropdown.
Conversation history
Today, prior turns are sent as a single concatenated input string with role labels (User: ... / Assistant: ...). This is the lowest-common-denominator shape that every runtime understands.
Future enhancement: structured history via metadata.conversation_history for runtimes that support it. Track this via issue #177.
Session reset
Switching agents in the dropdown clears the session and message history — session_id from agent A is meaningless to agent B. The Trash icon does the same on demand.
Auth-token setup
Agent mode requires the agent's runtime auth token to live in your workspace secrets backend. After you deploy:
agentbreeder secret set <agent-name>/auth-token
# (paste the token at the prompt)The dashboard now resolves it automatically. See Secrets — Per-agent auth tokens for details on the backend lookup.
When to use which
| Question | Use |
|---|---|
| "Does this prompt produce the right output?" | Model mode |
| "Does my deployed agent actually work end-to-end (tools, MCP, guardrails)?" | Agent mode |
| "Will my CI eval suite pass?" | Either — but Agent mode tests the live runtime |
Related
/api/v1/agents/{id}/invoke— the proxy endpoint Agent mode uses- Secrets backend — how
AGENT_AUTH_TOKENis resolved - Evaluations — turning a Playground turn into a regression test