QA: Sub-Agents — Langroid

Prerequisites

• Demo is deployed and accessible at /demos/subagents on the dashboard host
• Agent backend is healthy (/api/health); OPENAI_API_KEY is set on Railway; the FastAPI agent server exposes POST /subagents (see src/agent_server.py)

Test Steps

1. Basic Functionality

• Navigate to /demos/subagents; verify the page renders within 3s with the delegation log on the left and the CopilotChat pane on the right
• Verify data-testid="delegation-log" is visible with header "Sub-agent delegations"
• Verify data-testid="delegation-count" shows 0 calls
• Verify the empty-state copy "Ask the supervisor to complete a task. Every sub-agent it calls will appear here." is rendered
• Verify the chat input placeholder is "Give the supervisor a task..."
• Verify all 3 suggestion pills are visible with verbatim titles: "Write a blog post", "Explain a topic", "Summarize a topic"

2. Feature-Specific Checks

Live delegation log (running -> completed)

• Click the "Write a blog post" suggestion (sends a multi-step request that should trigger research -> write -> critique)
• Within 5s verify data-testid="supervisor-running" ("Supervisor running" pill) appears in the log header
• Within 10s verify the first data-testid="delegation-entry" appears with the Research badge (🔎 Research) and running status; the result body should show "Sub-agent running…"
• Within 30s verify the entry flips to completed status and a bulleted list of facts is rendered in the result body
• Verify a second data-testid="delegation-entry" appears with the Writing badge (✍️ Writing), goes through running -> completed, and renders a 1-paragraph draft
• Verify a third data-testid="delegation-entry" appears with the Critique badge (🧐 Critique) and renders 2-3 critiques
• Verify data-testid="delegation-count" updates to 3 calls (or more if the supervisor delegates again)
• After the run finishes, verify data-testid="supervisor-running" is no longer rendered and the chat receives a brief final summary

Sequential chaining

• Send "Explain how large language models handle tool calling. Research, write a paragraph, then critique."
• Verify the delegations appear in order: Research, then Writing, then Critique (the supervisor passes the prior step's output through task)
• Verify each entry's Task: line references the user's topic and (for Writing/Critique) cites the prior step

Multi-message persistence

• After a completed run, send another task ("Summarize a topic …")
• Verify NEW delegation entries are appended to the log (existing entries from the prior turn remain visible, count grows)
• Reload the page; verify the delegation log resets to empty and data-testid="delegation-count" shows 0 calls

3. Error Handling

• Send a trivially-conversational message like "Hi"; verify the supervisor either responds in plain text without delegating (count stays at 0 calls) or delegates only once and finishes — no infinite loop
• Verify DevTools -> Console shows no uncaught errors during any flow above
• If the secondary LLM fails (e.g. quota exhausted), verify a failed delegation entry is rendered with red status and a brief error message in the result body — the supervisor still produces a final user-facing message

Expected Results

• Page loads within 3 seconds
• Each delegation transitions from running to completed (or failed) within ~30s
• Delegation log entries appear in submission order and preserve across multiple supervisor turns within a single run
• The supervisor returns a final natural-language summary after the last sub-agent completes
• No UI layout breaks, no uncaught console errors