AI Agents UI

The AI Agents section of the sidebar provides visibility into the agentic AI pipeline and the reasoning agent runtime.

Agent Overview (/agents)

A health dashboard polling each agent service's /healthz endpoint and displaying a status card for each:

Agent	Port
Observability Agent	8092
Analysis Agent	8093
Action Agent	8094
Feedback Agent	8095
Recommendation Agent	8096

Below the cards, an Event Flow section shows the pipeline order (Observability → Analysis → Action → Feedback) as a simple labeled row — this is a plain summary strip, not the interactive Event Flow Diagram used elsewhere in these docs. That richer, clickable diagram (with real exchange names and the feedback-loop wiring gap called out) lives on the Architecture Overview page, not here.

Agent Runs (/agents/runs)

The runs page lists all reasoning agent executions with:

• Agent type badge
• Cluster ID
• Prompt preview (truncated)
• Status badge with animated spinner for running agents
• Duration
• Trigger type (manual / auto)
• Started timestamp
• View button → navigates to the run detail page

Starting a New Run

Click New Run to open the modal:

1. Select Agent Type from the dropdown
2. Enter an optional Cluster ID
3. Write your Prompt — describe what you want the agent to investigate

Click Start Run. The API creates the run record, starts the agent in a background goroutine, and returns a run_id. The modal closes and the runs list refreshes.

# Equivalent API call
POST /api/agents/runtime/api/v1/runs
Content-Type: application/json

{
  "agent_type": "sre_orchestrator",
  "cluster_id": "prod-us-east",
  "prompt": "Investigate why memory usage has been climbing for the last 2 hours"
}

agent-runtime actually supports seven agent types (see Agentic Runtime for the full list, tools, and models), but this modal's dropdown currently offers only four: SRE Orchestrator, Security Auditor, Cost Optimizer, and Incident Responder. The other three — NodeOps, Load Test Analyst, and an undocumented-elsewhere App Advisor type — can only be triggered via a direct API call today, not from this UI.

Run Detail (/agents/runs/{id})

A live streaming viewer with a two-panel layout:

Left Panel — Reasoning Trace

Shows the agent's internal reasoning process as it unfolds:

• Thinking blocks (purple, collapsible) — the agent's internal reasoning when interleaved thinking is enabled (SRE Orchestrator, Incident Responder, and NodeOps use Claude Sonnet 4.6 with interleaved-thinking-2025-05-14)
• Tool calls (blue) — each tool the agent invokes, with the input parameters
• Tool results (green) — confirmation that the tool completed, with response time in milliseconds
• Tool errors (red) — when a tool call fails

Right Panel — Agent Output

The agent's final text response streams in real time as the model generates tokens.

Streaming Implementation

The page uses the browser's native EventSource API to connect to:

GET /api/agents/runtime/api/v1/runs/{id}/stream
Content-Type: text/event-stream

Each SSE message is a JSON-encoded SSEEvent:

{ "type": "thinking", "run_id": "abc-123", "payload": "Let me start by checking..." }
{ "type": "tool_call", "run_id": "abc-123", "payload": { "tool": "get_cluster_health", "input": "{}" } }
{ "type": "tool_result", "run_id": "abc-123", "payload": { "tool": "get_cluster_health", "duration_ms": 143 } }
{ "type": "text", "run_id": "abc-123", "payload": "The cluster health score is 72/100..." }
{ "type": "done", "run_id": "abc-123", "payload": null }

The stream closes on a done or error event. If the connection drops for any other reason, the page currently does not attempt to reconnect — it just marks the stream disconnected. Refreshing the page re-opens a new stream from the current point rather than resuming the old one; nothing is lost from the run itself (tool calls are persisted server-side), but you won't see a live reconnect happen automatically.

Individual Agent Pages

Sub-pages exist today for three of the five pipeline agents — /agents/observability, /agents/analysis, and /agents/recommendations — each showing recent telemetry snapshots, analysis results, or recommendations fetched live from the corresponding service. /agents/actions is linked from the sidebar's AI Agents menu but has no page built behind it yet; there's no Feedback Agent page or sidebar link at all today.