Building AI Agents That Actually Work: Introducing Nova Act as a Service
Source: https://www.amazon.science/blog/introducing-nova-act-as-a-service Author: AGI Lab (Amazon) Date: 2025-12-03
Summary
Amazon’s announcement of Nova Act as a Service — a computer-use agent API that performs web browser actions. Positions against Anthropic’s computer use and OpenAI’s operator. Key differentiator: reliability focus over capability ceiling.
Key Claims
- Nova Act is designed for production reliability, not benchmark performance — explicitly trades peak capability for consistency.
- Key design choice: modular action primitives (navigate, click, type, read) vs. end-to-end task completion — easier to debug and recover from errors.
- Rate-limits and retries are first-class citizens in the API — acknowledges that agents need robust error handling.
- Reliability metric: they measure “task success rate” across many runs, not single-shot performance.
- Observation: most computer-use failures happen at the “last mile” — the model understands the task but fumbles a specific UI interaction.
Entities
- Amazon / AWS — Nova Act product
- Anthropic — competitor in computer use
Concepts
- Coding Agents — CUA is adjacent to coding agents
- Reward Hacking — reliability focus avoids gaming benchmarks