DEVOURED
Can Agents Use a Computer Yet? We've Got the Data
Computer-using agents have reached a production milestone, effectively managing repetitive back-office tasks by navigating legacy software interfaces without needing custom APIs.
What: Computer-using agents (CUAs) currently achieve an 85% success rate on the OSWorld-Verified benchmark, up from 42% in 2025. These agents are being deployed in production for standardized workflows like IT ticket processing in ServiceNow and CRM record updates at a cost of $6-8 per hour, making them economically competitive with offshore BPO labor.
Why it matters: The industry is shifting from evaluating AI on model-layer benchmarks to prioritizing infrastructure-layer concerns like reliability, error handling, and business process integration, effectively commoditizing the underlying frontier models.
Takeaway: Focus on high-volume, standardized tasks where failure modes are observable and can be mitigated by deterministic 'harnesses' that catch errors before they impact downstream systems.
Deep dive
- Computer-using agents are primarily effective for 'protocol-following' tasks where a clear, repeatable path exists.
- Production systems often use a 'caching' strategy: an agent handles a workflow once, which is then converted into deterministic code for subsequent runs, with the agent re-engaged only for recovery or error handling.
- The primary technical bottleneck has shifted from model intelligence to the orchestration, verification, and error-handling layer.
- Cost-effective deployments use a mix of frontier models for reasoning and cheaper, smaller models for routine UI interactions.
- Business models for CUA vendors vary, but pricing per outcome or task is becoming common to manage unpredictable inference costs.
- High-value deployments include self-healing scrapers, automated data entry for legacy portals, and IT ticket triage.
Decoder
- BPO (Business Process Outsourcing): Contracting non-primary business processes, such as data entry or customer support, to third-party providers, typically in regions with lower labor costs.
- Harness: A software wrapper or orchestration layer that provides an agent with access to a specific environment (like a VM or browser), handles error logging, and manages retry logic.
- System of Record: A central software application (like a CRM or ERP) that acts as the authoritative source for data within an organization.
- OSWorld-Verified: A standard benchmark that evaluates how well AI agents can operate across different desktop environments (macOS, Windows, Ubuntu) to complete specific user tasks.
Original article
Full article content is not available for inline reading.
Explore all 43 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi