Skip to content

HUMAN_IN_LOOP.md

Latest commit

 

History

History
121 lines (82 loc) · 2.25 KB

HUMAN_IN_LOOP.md

File metadata and controls

121 lines (82 loc) · 2.25 KB

Human-in-the-Loop - The Critical Element

Agent Cube is NOT fully automated - it's human-assisted automation

🎯 The Reality

Agent Cube tells you when it needs help:

Common Interventions

1. Judge Doesn't Save Decision (~30% with Gemini)

cube decide task
→ "Missing: Judge 3"

cube resume judge-3 task "Write your decision JSON"
→ Judge files decision

cube decide task  
→ "3/3 decisions, Winner: B"

2. Synthesis Needs Clarification

cube auto task --resume-from 6
→ Prompter generates synthesis
→ Winner addresses blockers

cube status task
→ Shows: "Peer review in progress"

3. All Judges Approved Wrong Architecture

cube decide task
→ "3/3 APPROVED, Winner: B"

[Human reviews PR]
→ "Wait, this isn't following the architecture!"

[Close PR, restart with clearer requirements]

4. Network Error During Panel

Judges running...
→ "cursor-agent network error"

cube auto task --resume-from 4
→ Re-runs panel
→ Completes successfully

💡 The Tool Guides You

Agent Cube shows exactly what to do:

⚠️  Missing peer review decisions (2/3)

Options:
  1. Get missing judge:
     cube resume judge-3 task "Write decision"
  
  2. Continue with 2/3:
     cube auto task --resume-from 8

No guessing. Clear next steps.

🔄 Continuous Improvement

The goal:

  • Catch more edge cases
  • Better error recovery
  • Smarter defaults
  • Less human intervention

But always:

  • Human validates architectural decisions
  • Human reviews final PR
  • Human decides when to merge

Future:

  • Auto-orchestration (run multiple tasks in dependency order)
  • Learning system (which prompts work best)
  • Better Gemini integration
  • Smarter synthesis

🎓 FOR THE SESSION

Key message:

"Agent Cube requires human guidance.

But it tells you EXACTLY when and how.

Judges don't save? Here's the command. Need synthesis? Here's what to do. Something breaks? Here's the recovery.

It's not magic. It's robust tooling with clear feedback.

The goal: Reduce intervention over time. Today: ~5 interventions per complex feature. Future: ~1-2 interventions.

But humans always validate the final result."

ADD THIS TO SLIDE 15 (LIMITATIONS & FUTURE)