Scale computer-use 2.0 with open-source drivers, cross-OS sandboxes, and benchmarks for training, evaluation, and data generation.
Rust agentai-agentapplecomputer-usecomputer-use-agentcontainerizationcuadesktop-automationhacktoberfestlumemacosmanusoperatorswiftvirtualizationvirtualization-frameworkwindowswindows-sandbox
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
- Python loops over items: never the agent
- Two agents per item: creator does work, auditor verifies independently
- One submit tool per agent: no get_next, no submit_results, no update_and_submit
- Auditor uses a checklist: returns
[{item, passed: bool}], never a score - Python calculates scores: from checklist bools, never ask agent for a number
- Session continuation on retry: pass
session_idso creator keeps context - Pre-boot sandbox in Python: agents never waste turns on setup
- 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