Skip to content

subagents.md

Latest commit

 

History

History
60 lines (42 loc) · 3.73 KB

subagents.md

File metadata and controls

60 lines (42 loc) · 3.73 KB
Raw

QA: Sub-Agents — Claude Agent SDK (Python)

Prerequisites

  • Demo is deployed and accessible at /demos/subagents on the dashboard host
  • Agent backend is healthy (/api/health); ANTHROPIC_API_KEY is set on Railway; the FastAPI backend exposes POST /subagents

Test Steps

1. Basic Functionality

  • Navigate to /demos/subagents; verify the page renders within 3s with a left-side delegation log and a right-side CopilotChat pane
  • Verify data-testid="delegation-log" is visible with heading "Sub-agent delegations"
  • Verify data-testid="delegation-count" reads "0 calls" on first load
  • Verify the empty-state placeholder reads "Ask the supervisor to complete a task. Every sub-agent it calls will appear here."
  • 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

Multi-stage Delegation Flow

  • Click the "Write a blog post" suggestion (cold-exposure training prompt)
  • Within 5s verify data-testid="supervisor-running" appears with the "Supervisor running" pulse indicator
  • Within 30s verify at least one data-testid="delegation-entry" appears with badge "Research" and data-testid="delegation-status" initially reading running
  • Verify the entry's status flips to completed once the sub-agent returns and the result text is visible inside the entry's white inner panel
  • Within 60s total verify additional delegation entries appear in order: Research -> Writing -> Critique (3 entries total in most cases)
  • Verify data-testid="delegation-count" updates to match the number of entries (e.g. "3 calls")
  • Verify the supervisor's final chat reply includes a brief summary of the produced deliverable

Delegation Entry Layout

  • Each data-testid="delegation-entry" shows: a #N index, a sub-agent badge with the correct emoji (🔎 Research / ✍️ Writing / 🧐 Critique), a status chip, the task text after "Task:", and the sub-agent's result rendered with whitespace preserved
  • Hover the supervisor running chip while a delegation is in flight — verify the pulse animation is present (no static-only state)

Sub-Agent Independence

  • Click "Explain a topic" (LLM tool calling prompt) and wait for completion
  • Verify the writing entry's result is a single polished paragraph (the writing sub-agent's signature)
  • Verify the research entry's result is a bulleted list of 3-5 facts (the research sub-agent's signature)
  • Verify the critique entry's result contains 2-3 actionable critiques

Supervisor State Reset Across Tasks

  • After the first run completes, click "Summarize a topic" (reusable rockets)
  • Verify NEW delegation entries are appended to the existing list (count keeps growing) — confirms state["delegations"] accumulates across turns within the same thread

3. Error Handling

  • Send an empty message; verify it is a no-op
  • If a sub-agent call fails (e.g. due to upstream rate limit), verify the failing entry is rendered with status failed and a result line starting with "sub-agent call failed:" — confirms the fail-loud path
  • Verify DevTools -> Console shows no uncaught errors during any flow above

Expected Results

  • Page loads within 3 seconds
  • First delegation entry appears within 30 seconds of submitting a non-trivial task
  • Each delegation entry transitions from running -> completed (or failed) and the count badge stays in sync with state["delegations"].length
  • Supervisor's final chat reply summarises the work and arrives within 90 seconds of submission
  • No UI layout breaks, no uncaught console errors