external-ai-agents

--- name: external-ai-agents description: Use when stuck in loops, need fresh perspective, or want to simulate multiple viewpoints - invokes cursor agent or gemini CLI as external subagents for model diversity and outside opinions --- # External AI Agents Invoke Cursor or Gemini CLI tools as subagents for fresh perspectives, breaking loops, or multi-perspective simulation. ## When to Use - **Stuck in a loop** - same approach failing repeatedly, need outside view - **Multi-perspective review** - simulate architect vs implementer, security vs velocity - **Model diversity** - different training = different blind spots - **Sanity check** - "am I overcomplicating this?" ## Invocation Patterns ### Basic (small output expected) ```bash cursor agent --prompt "Review this approach: ..." gemini --prompt "What am I missing here: ..." ``` ### File-based (large output, searchable) ```bash cursor agent --prompt "..." > /tmp/cursor-review-$(date +%s).md gemini --prompt "..." > /tmp/gemini-analysis-$(date +%s).md # Then grep for key insights grep -i "problem\|issue\|suggest\|instead" /tmp/cursor-review-*.md ``` ### With context file ```bash cursor agent --prompt "Review this code for issues: $(cat src/worker.rb)" gemini --prompt "$(cat PLAN.md)" --prompt "What holes are in this plan?" ``` ## Integration with Orchestration ### With subtask (parallel perspectives) ```bash # Get two perspectives on same problem subtask ask "Review the auth implementation in src/auth/ for security issues" & cursor agent --prompt "Review src/auth/ for security issues" > /tmp/cursor-auth.md & wait # Compare findings diff <(grep -i "issue\|vuln" /tmp/cursor-auth.md) \ <(subtask show auth-review | grep -i "issue\|vuln") ``` ### With Task tool (Claude orchestrates) Dispatch external agent, then synthesize: ```bash # In background gemini --prompt "Analyze the performance bottleneck in $(cat profile.json)" > /tmp/gemini-perf.md # Claude reads and synthesizes # Use Read tool on /tmp/gemini-perf.md, combine with own analysis ``` ## Multi-Perspective Simulation ### Architect vs Implementer ```bash cursor agent --prompt "As a senior architect: Is this design over-engineered? $(cat DESIGN.md)" > /tmp/architect.md gemini --prompt "As an implementer who has to build this: What's impractical? $(cat DESIGN.md)" > /tmp/implementer.md ``` ### Red Team / Blue Team ```bash gemini --prompt "Find security holes in: $(cat src/api/auth.rb)" > /tmp/red-team.md cursor agent --prompt "How would you defend against: $(cat /tmp/red-team.md)" > /tmp/blue-team.md ``` ### Devil's Advocate ```bash # When you're convinced of an approach gemini --prompt "Argue against this decision. Be harsh: We chose PostgreSQL over MongoDB because..." ``` ## Breaking Loops Pattern When stuck after 3+ similar attempts: 1. **Capture context**: What you've tried, what's failing 2. **Ask for reframe**: "What am I missing?" not "How do I fix X?" 3. **File output**: Large responses need grep filtering ```bash cat << 'EOF' > /tmp/stuck-context.md ## What I'm trying to do [goal] ## What I've tried 1. [approach 1] - failed because [reason] 2. [approach 2] - failed because [reason] ## Current error/blocker [details] EOF gemini --prompt "$(cat /tmp/stuck-context.md)" --prompt "What assumption am I making that might be wrong?" > /tmp/fresh-perspective.md # Extract actionable suggestions grep -A2 -i "try\|suggest\|instead\|consider" /tmp/fresh-perspective.md ``` ## Output Management | Scenario | Approach | |----------|----------| | Quick sanity check | Direct stdout | | Detailed review | File + grep | | Comparison | Multiple files + diff | | Synthesis needed | File, then Read tool | ## Common Mistakes - **Too much context** - trim to essentials, models get lost in noise - **Asking "how to fix"** - ask "what am I missing" for better reframes - **Ignoring output** - file it, grep it, don't let insights vanish - **Single perspective** - the value is in *multiple* viewpoints