Agent Playground.
The Agent Playground is a high-fidelity sandbox environment designed for iterating on agent formulas. Test vision, voice, and tool execution in a safe, isolated context before deploying to production.
01. Isolated Sandbox Environment
Playground interactions are strictly partitioned from your production environment. This allows you to test destructive tool calls or experimental prompts without impacting your long-term agent memory or production data.
- Episodic-Only Memory: Interactions are not stored in the agent's long-term vector mesh, preventing "hallucination poisoning" during testing.
- Formula Snapshots: Test multiple versions of a system prompt side-by-side to compare reasoning accuracy.
02. Multi-Modal Perception Testing
The playground is built for the modern AI stack, supporting native testing for vision and audio interactions.
Vision Sandbox
Upload or paste images directly into the chat to verify your agent's OCR and visual reasoning capabilities.
Voice Bridge
Toggle "Live Voice" to test STT/TTS settings in real-time via WebSockets. Hear your agent's persona before going live.
03. Real-Time Trace & Diagnostics
Unlike standard chat interfaces, the playground provides a transparent window into the agent's internal reasoning cycle.
- Orchestration Trace: See exactly when the agent is "Thinking", "Transcribing", or "Synthesizing".
- Tool Output Inspection: View the raw JSON output of every tool call to debug parameter mapping and integration errors.
- Latency Monitoring: Track the millisecond duration of every turn to optimize for the best user experience.
04. Live Voice Bridge
Directly test your voice providers (ElevenLabs, Whisper, etc.) without writing a single line of frontend code. The Playground uses the same /v1/hub/voice/stream WebSocket API that your production app will use, ensuring that what you hear in the sandbox is exactly what your users will hear.