Streaming & Real-Time

MeetLoyd provides comprehensive real-time streaming capabilities for agent responses, event monitoring, and live collaboration. All streaming uses Server-Sent Events (SSE) for efficient, browser-compatible real-time communication.

Conversation Streaming

Stream agent responses token-by-token for real-time chat experiences. As the agent thinks, you see each word appear -- just like typing in a messaging app.

Stream Event Types

Event	Description
token	A generated text token (partial response)
tool_call	Agent started using a tool
done	Response complete (includes token usage and cost)
error	An error occurred
thinking	Agent is reasoning (for models with extended thinking)
heartbeat	Keep-alive signal (every 5 seconds)

Heartbeat

MeetLoyd sends heartbeat events every 5 seconds during streaming to prevent reverse-proxy idle timeouts. Your client should ignore these events -- they carry no data. If you stop receiving heartbeats, the connection has likely dropped.

Live Message Injection

You can send messages to an agent while it is still streaming a response. This is useful for corrections, additional context, or changing direction mid-task.

How It Works

1. Your message is stored in an in-memory queue keyed by conversation or thread ID
2. At the top of each executor iteration (between tool calls), the queue is drained
3. Messages are appended to the conversation history with a clear "live message" marker
4. The LLM sees them as new context and can acknowledge, pivot, or continue as needed
5. Messages auto-expire after 10 minutes if never drained

info

Injection only works while the agent is actively streaming. If the stream has finished, send a new message normally.

Command Center Streaming

Real-time monitoring of agent executions across your organization. The command center streams events for all running agents, giving you a live operational view.

Event Types

Event	Description
Run started	Agent execution initiated
Run status changed	Execution state transition
Run progress	Token and cost progress updates
Tool started/completed	Tool execution lifecycle
LLM call	LLM API call metrics
Memory operation	Memory read/write events
Error	Execution error
Alert triggered/resolved	Alert condition changes

You can filter the stream by specific agents or event types to focus on what matters.

Setting Up Streaming

Conversation Streaming

Conversation streaming is enabled by default. When sending a message through the dashboard or the AG-UI endpoint, responses stream automatically. The dashboard handles reconnection and buffering for you.

For custom integrations, connect to the SSE endpoint and handle events by type. Key implementation considerations:

• Handle reconnection -- Use exponential backoff if the connection drops (max 5 attempts, up to 30s delay)
• Clean up connections -- Always close the EventSource when navigating away or when the response completes
• Buffer UI updates -- Batch token events using requestAnimationFrame for smoother rendering
• Ignore heartbeats -- Heartbeat events carry no data; they exist solely to keep the connection alive

Command Center

The command center stream is available at the admin level. Configure which agents and event types you want to monitor, and the stream delivers matching events in real-time.

AG-UI Streaming

The AG-UI endpoint provides team-level streaming for the Openspace interface. It supports:

• Sending messages and receiving streamed responses
• Live message injection during active streams
• Real-time activity feeds per team
• 5-second heartbeats to survive proxy idle timeouts

Data Streaming Pipelines

Build real-time data processing pipelines for continuous data flows. Pipelines connect a source (agent events, webhooks, database changes) to a destination (analytics, webhooks, notifications) with optional transforms in between.

Pipeline Components

Component	Options
Sources	Agent execution events, webhooks, database changes, API events, scheduled triggers
Transforms	Filter, map, aggregate, window, batch
Destinations	Webhook, database, analytics (Mixpanel/Amplitude/Segment), notifications (Slack)

Transform Operations

Transform	Description
Filter	Keep only events matching a condition
Map	Reshape event data (rename fields, compute values)
Aggregate	Group and summarize (sum, average, count)
Window	Time-based aggregation (tumbling, sliding, session windows)
Batch	Collect events into batches by size or time

Managing Pipelines

From the dashboard, you can start, pause, resume, and stop pipelines. Each pipeline tracks metrics including items processed, failures, throughput, and latency. You can also create checkpoints for recovery.

Best Practices

Handle Reconnection Gracefully

SSE connections can drop due to network issues. Implement exponential backoff reconnection so the user experience remains smooth.

Clean Up Connections

Always close EventSource connections when they are no longer needed. Leaked connections waste resources on both client and server.

Buffer UI Updates

Rendering every single token event can cause jank. Batch updates using requestAnimationFrame for a smoother experience.

Monitor Pipeline Backpressure

If a pipeline's throughput cannot keep up with incoming events, backpressure builds. Watch pipeline metrics and consider increasing batch sizes or scaling destination capacity.

---

Next: Learn about Avatars for video responses from agents.