Skip to main content

Testing

The streaming test window is the panel behind the Test button on agent and manager pages. It runs your agent live over a streaming connection and shows everything as it happens — text as it's generated, every tool call, each sub-agent delegation on a manager, and audio responses — so you can see not just what an agent answered, but how it got there.

What it's for​

An agent that "mostly works" usually fails somewhere specific: a tool called with the wrong argument, a sub-agent given a vague instruction, a critic loop rejecting good output. The test window exposes those internals for a real run against your current draft — no publishing, no separate harness. Use it after every meaningful change to instructions, tools, or template variables, before traffic (or the version you publish) depends on the behavior.

Open an agent or manager in the editor and click the purple Test button (next to Real Time Voice and Save). A panel slides in from the right — titled Autobot (Streaming) for agents, Test Multi-Agent (Streaming) for managers — and connects immediately; the colored dot under the title shows the connection status and the current conversation ID. The refresh icon in the header resets everything and reconnects without closing the window; closing it discards the session.

Key concepts​

  • Variables input — if the agent or manager defines prompt template variables, a Prompt Template Variables section appears above the input with one field per variable (its description as the placeholder). Values are sent with each message, so you can test how different substitutions change behavior.
  • Text or audio output — a Text / Audio toggle above the input selects the response mode. Audio responses render as playable cards ("Audio Response") that fill as chunks stream in, and an Audio Transcription History section collects the transcripts.
  • Voice input — the microphone button records 16 kHz audio (with a live duration counter) and attaches it to your message, so you can test how a voice-driven agent handles real speech. A paperclip menu also attaches images and documents where the model supports them.
  • A conversation, not one-shots — messages accumulate in a single conversation until you hit refresh, so multi-turn behavior (memory, follow-ups, clarifying questions) is testable. There's no stop button; the input simply unlocks when the response completes.

How to read the stream​

Each run unfolds as a sequence of message types in the transcript:

  • Your message — right-aligned, with chips for any recorded audio, images, or files attached.
  • Processing indicator — bouncing dots ("Agent is processing...") while the agent works before content arrives.
  • Tool activity — collapsible cards grouped per tool invocation, headed "<tool name> (N messages)" with a blue gear icon. Collapsed, they show the latest status line; expanded, every status message with its variables and timestamp. Groups auto-collapse when the run finishes — expand them when a result looks wrong to see exactly what the tool received and reported.
  • Sub-agent delegations (manager window only) — each delegation is a card named for the sub-agent, showing the instruction the manager gave it, its status, and a nested Tools Used list with that sub-agent's own tool invocations. This hierarchy — manager → sub-agent → tools — is the main thing to study when a multi-agent run goes sideways: bad final answers are usually bad instructions.
  • Critic attempts — agents with a critic loop show attempt cards: Initial Attempt, Retry N, and a final Approved (green) or Rejected (red) verdict with the critic's reasoning. Rejected attempts auto-collapse; the approved one stays open.
  • Streaming text — the reply renders token by token with a cursor until complete.
  • Errors — a red Streaming Error card with the raw message; connection problems also turn the status dot red while the window retries.

How to test an agent​

  1. Open the agent, make your edits, and click Save, then Test.
  2. Fill in any Prompt Template Variables and pick Text or Audio output.
  3. Send a realistic message (Enter to send, Shift+Enter for a new line) — the phrasing a real user would use, not a happy-path prompt.
  4. Watch the tool cards as much as the answer: verify each tool fired when expected, with the arguments you intended.
  5. Continue the conversation to probe follow-ups and edge cases; click the header's refresh button to reset and try a clean variation.
  6. For managers, expand each sub-agent card and check the instruction the manager wrote — then tighten the manager's instructions if delegations are vague.

Test conversations are real runs — tools execute with real side effects (emails send, rows write). Point destructive tools at test data first.