Agentic AI

The trycua/cua repository provides infrastructure to scale computer-use 2.0 AI agents, featuring open-source drivers for cross-OS desktop automation. It includes Cua Spaces for managing full desktop environments on Mac and owned machines, alongside Lume for local VM creation. The project also offers CUA-S1 specialized decision models and Cua Bench for evaluating agent performance and generating training data.

• Cua Spaces enables agents to operate full desktop environments on Mac and owned hardware.
• Cua Driver automates desktop interactions, allowing verification of actions like calculator results.
• Cua Bench provides tools for evaluating computer-use agents and generating training data.
• CUA-S1 offers small, specialized models tailored for computer-use decision-making tasks.
• Lume facilitates creating Tahoe VMs with SSH connectivity for local sandboxing.

Researchers evaluated nine models ranging from 0.8B parameters to frontier-tier LLMs across five distinct call sites in a deployed home-automation system. Using unmodified production prompts and real-world data, the study reveals that model capability is not uniformly ordered across different task types. The findings challenge the industry practice of using a single, large model for all agentic operations, demonstrating that smaller models often suffice or even excel at specific, less complex call sites.

• Capability ranking varies by task; larger models do not universally outperform smaller ones.
• Agentic systems have heterogeneous call sites with varying difficulty levels.
• Smaller models can handle specific tasks more efficiently than frontier-tier LLMs.
• A single large model strategy is inefficient for systems with diverse call requirements.

This paper introduces Humanize, a multi-agent workflow designed to solve the reliability gap in agentic coding where agents struggle to judge their own completion. The system uses explicit judgement engineering with deterministic hooks that route work between planning, implementation, and review roles. Key to this approach is the enforcement of 72 mechanical gates, including mandatory human approval for plan contracts and independent vendor-based reviewer agents for completion decisions. The process is modeled as a Markov chain over repository states to ensure rigorous defect detection through alternating builder and reviewer sampling.

• Relies on judgement engineering rather than model capability to ensure reliable code completion.
• Enforces 72 mechanical gates via deterministic hooks to control workflow boundaries.
• Requires human approval for plan contracts before any implementation begins.
• Uses an independent reviewer agent from a different vendor to decide completion status.
• Models the workflow as a Markov chain to manage defect survival probabilities.

Researchers address how persistent LLM memory retains rejected hypotheses as facts, creating durability for false reasoning. They introduce 'discourse ownership', a signal representing the context or speaker licensing a proposition. This approach aims to detach propositions from their original validity scope before committing them to long-term storage.

• Persistent memory risks turning simulated or rejected agent plans into durable false facts
• Discourse ownership captures the context that licenses a proposition's validity
• LLMs already contain causal signals for ownership that can be leveraged for filtering
• Scope-aware rules prevent context-less propositions from contaminating long-term memory

Research highlights how supervisory governors can inadvertently cause persistent task failures by interfering with regulated agents. In controlled file-recovery tests, cost-blind governors triggered blocking when regulatory intensity increased and tool failures occurred. The study compares this against an adaptive backoff mechanism that lowers intervention probability based on recent failure history, showing improved completion rates.

• Governors can become the primary source of blocking if they ignore cost implications of interventions.
• Adaptive backoff using moving averages of induced failures outperforms fixed weak governors.
• Persistent blocking occurs even with simple stochastic-policy agents under result replacement attacks.
• Implementing failure-aware backoff rules is critical for resilient tool-using agent architectures.