Skip to content

armgit / cua

Public Imported from GitHub
0

Scale computer-use 2.0 with open-source drivers, cross-OS sandboxes, and benchmarks for training, evaluation, and data generation.

1.7 KB
Raw Blame

Agent Design

The Pattern

for item in items:
    session_id = None
    for attempt in range(3):
        # Creator agent: sandbox + tools + ONE submit tool
        creator_result = await run_agent(item, session_id, creator_tools)
        session_id = creator_result.session_id

        # Auditor agent: sandbox + tools + ONE submit_audit tool + checklist
        audit = await run_agent(creator_result.submission, auditor_tools)

        # Python: extract bools, calculate score, decide
        score = sum(c["passed"] for c in audit.checklist) / len(audit.checklist)
        if score >= threshold:
            break
        # else: creator retries with session continuation + auditor feedback

Rules

  1. Python loops over items: never the agent
  2. Two agents per item: creator does work, auditor verifies independently
  3. One submit tool per agent: no get_next, no submit_results, no update_and_submit
  4. Auditor uses a checklist: returns [{item, passed: bool}], never a score
  5. Python calculates scores: from checklist bools, never ask agent for a number
  6. Session continuation on retry: pass session_id so creator keeps context
  7. Pre-boot sandbox in Python: agents never waste turns on setup
  8. Action-script prompts: numbered steps with exact tool calls, not open-ended

Anti-Patterns

  • Agent loops over items → runs out of turns after 1-2
  • Agent self-verifies → self-confirmation bias
  • Multiple submit tools → agent gets confused
  • Auditor gives scores → arbitrary, unreproducible
  • Fresh agent on retry → loses context
  • Agent boots sandbox → wastes turns
  • Open-ended prompts → agent explores instead of executing