Open-source AI agent 'AutoBot' for Mac released, outperforms OpenAI Sol Max on OSWorld

The Autonomous Production team has released AutoBot, an open-source agentic harness that allows real-time voice control of long-running AI tasks in Mac environments. This tool integrates with Native ChatGPT and is distributed under the MIT license.

Benchmark Performance

AutoBot demonstrated superior performance compared to competing models on major agent benchmarks.

• OSWorld 2.0: Achieved a Binary Accuracy of 32.41% and Partial Accuracy of 64.28% across 108 tasks. This represents a 18.5% higher task completion rate than the OpenAI Sol Max baseline and surpasses Anthropic Claude Opus 5 Max.
• AssistantBench: Ranked first on the hidden-test leaderboard with 50.70% (181 tasks).

Core Technologies and Differentiators

The key mechanisms driving performance improvements are the self-repairing harness, hierarchical disk memory (nightly consolidation), persistent task graphs, and local CPU orchestration that reduces inference costs. It offers the following operational differentiators compared to existing tools:

• Memory boundary management: Separates personal, family, work, and shared zones, and specifies reuse rules.
• Result tracking and verification: Manages incomplete items through Owner/Destination/Completion assignments, and separates validator labels that check accounts and content before writing.
• 5-step completion process: Proceeds through Research → Draft → Destination → Readback → Completion stages, blocking writes upon verification failure.

Operational Limitations and Security

AutoBot is optimized for single-user Mac environments and does not guarantee success for multi-tenant services or third-party tasks. Regarding security, Privacy zones and validator separation are based on procedural rules rather than OS-level sandboxes, so separate device authentication (MFA) is essential for sensitive tasks. Since not all requests may remain local, it is necessary to check ChatGPT data control settings.