Anthropic hardens Claude for production-grade computer-use agents

Anthropic spent March focused less on headline model numbers and more on making Claude behave like a dependable production component in agentic stacks. The biggest change for builders is the improvement in its computer-use capabilities: error rates on desktop application interactions reportedly dropped by around 40%, with stronger handling of dynamic UI elements, modal dialogs, and multi-step forms. That directly expands the range of enterprise workflows — particularly back-office RPA tasks — that can be automated end-to-end.

What changed. Claude’s computer-use APIs and underlying behaviors were upgraded to handle more complex, stateful desktop workflows while failing less often.

Complementing this, Anthropic introduced new streaming and batching API endpoints targeted at high-throughput scenarios typical in agentic deployments. Instead of firing off thousands of single requests, teams can send large batches and stream responses, significantly reducing latency and cost for pipelines like content generation, analysis, and multi-agent coordination. Combined, these changes signal a push to make Claude a central node in production automation rather than just an interactive assistant.

Why it matters. Desktop automation agents live or die on reliability and throughput; lower interaction error rates and cheaper batching are exactly what’s needed for credible, scalable pilots.

Builder takeaway. Revisit any workflows you previously deemed too fragile for Claude-driven computer-use—these upgrades may cross the reliability threshold needed for real RPA deployments, especially when paired with the new streaming/batching APIs.