Bridge Delegation

Delegation is the most powerful bridge capability. It lets one workspace invoke another workspace's AI agent, have it process a task using its own tools and data connections, and return the result — all within a single conversation flow.

How Delegation Works

When a user triggers a delegation, the following sequence occurs:

1. User sends message     →  "@ai ask analytics to summarize last week's revenue"
2. Source AI recognizes    →  Calls bridge_workspace tool with target ID + task
3. Control plane relay     →  HMAC-signed request forwarded to target workspace
4. Target AI processes     →  Uses its own provider, model, tools, and data connections
5. Result returned         →  Response relayed back through control plane
6. Source AI presents      →  Formats and displays the result in the source chat

The entire flow happens within the source workspace's conversation. The user sees the delegated response appear inline, as if the source AI answered — but the work was done by the target workspace's AI.

In-Chat Experience

When a delegation is in progress, the source workspace's chat shows a streaming delegation response. You'll see:

1. An indicator that work has been delegated to the target workspace.
2. The response streams in as the target AI generates it.
3. The final result appears in the conversation, attributed to the bridge.

This keeps the user experience seamless — no need to switch workspaces or check separate threads.

Example

A user in the Engineering workspace types:

@ai ask the analytics workspace to summarize last week's revenue by product line

The Engineering workspace's AI:

1. Identifies this as a delegation request.
2. Calls the bridge_workspace tool, targeting the Analytics workspace.
3. Sends the task: "Summarize last week's revenue by product line."

The Analytics workspace's AI:

1. Receives the delegated task.
2. Runs a query_bigquery tool call against its connected revenue dataset.
3. Summarizes the results.
4. Returns the response.

The Engineering workspace displays the Analytics AI's summary directly in the chat.

Timeout

Delegated tasks have a 120-second timeout. If the target workspace's AI does not return a result within 2 minutes, the delegation fails and the source workspace receives a timeout error.

:::tip Handling Complex Tasks For tasks that may take longer than 120 seconds, consider breaking them into smaller sub-tasks or using the message mode to trigger async processing on the target workspace. :::

Auto-Wake

If the target workspace is sleeping (scaled down due to inactivity), Roundtable automatically starts it when a delegation request arrives.

• The cold-start time is included in the 120-second timeout window.
• Typical cold-start takes 10–30 seconds depending on the workspace configuration.
• Subsequent delegations to an already-running workspace are much faster.

:::info No Manual Intervention You don't need to manually start a sleeping workspace before delegating to it. Roundtable handles wake-up automatically. Just send the request. :::

Retry Logic

If a delegation request fails due to a transient error (network issue, temporary target unavailability), Roundtable retries automatically:

Attempt	Delay
1st retry	~1 second
2nd retry	~2 seconds
3rd retry	~4 seconds

• 3 retries with exponential backoff.
• If all retries fail, the source workspace receives an error message.
• Permanent errors (invalid bridge, permission denied, target deleted) are not retried.

Best Practices

1. Write Descriptive System Prompts for Target Workspaces

The target workspace's AI responds to delegated tasks using its system prompt as context. A vague system prompt leads to vague responses.

❌ Poor system prompt:

You are a helpful assistant.

✅ Effective system prompt:

You are the Analytics AI for Acme Corp. You have access to BigQuery datasets
containing revenue, customer, and product data. When asked for summaries,
always include date ranges, totals, and percentage changes. Format financial
figures with currency symbols and two decimal places.

2. Scope Tools and Data Connections

Ensure the target workspace has the right tools enabled and data connections configured for the tasks it will receive. A workspace without query_bigquery can't answer data questions, no matter how good its system prompt is.

3. Test Delegations with Simple Tasks First

Before relying on complex cross-workspace workflows, test with a basic delegation:

@ai ask the [target] workspace to describe what it can do

This confirms the bridge is active, permissions are correct, and the target AI is responsive.

4. Keep Delegated Tasks Focused

Delegation works best when tasks are specific and self-contained. Instead of:

@ai ask analytics to analyze everything about our business

Try:

@ai ask analytics to compare Q1 vs Q2 revenue for the enterprise segment

5. Consider Timeout Constraints

Design delegated tasks to complete within the 120-second window. If you need a long-running analysis, have the target workspace write results to a shared data store and use a message bridge to notify when done.

Troubleshooting

Issue	Cause	Solution
"Bridge not found"	Bridge was deleted or never created	Create a new bridge in the Bridges panel
"Permission denied"	Source workspace lacks delegate permission	Update bridge permissions to include Delegate
"Timeout"	Target took longer than 120s	Simplify the task or break into sub-tasks
"Target workspace unavailable"	Target workspace was deleted	Verify the target workspace still exists
No response	Target AI has no relevant tools	Add required data connections and tools to the target workspace