计算机代理的下一步：从自我排错到异步工具、决策 API 与长期上下文

A personal cloud computer expands what an agent can execute across desktop and web software, but the value of that delegation still depends on the user's actual tasks.

The important improvement is not merely starting a task, but debugging failures and retrying effectively as both the model and its execution harness become more capable.

Accessibility information and direct DOM access can reveal context that screenshots omit, allowing an agent to inspect more of an interface and reduce unnecessary interaction steps.

As model actions become faster, external loading and synchronization delays can dominate task latency, making efficient event detection as important as faster inference.

A broadly capable computer agent needs consent for consequential actions and scoped access, while interactive testing can help verify the software it creates before delivery.

Asynchronous tool calling and mid-turn steering let a model continue reasoning while tools run and incorporate new input, with bidirectional communication supporting that interaction.

The first Decisions API implementation constrains outputs and optimizes parallel low-latency classification on existing Luna weights rather than introducing a newly trained general reasoning model.

Low latency and structured outputs do not establish calibrated confidence, so decision systems still need task-specific accuracy checks and an explicit policy for uncertain cases.

Cache-aware applications can reduce repeated prompt cost through reuse and pre-warming, but longer retention guarantees described for a preview customer should not be assumed universally available.

Automatic and manually triggered compaction offer different control over long-running context, while higher-level agent APIs must balance useful abstraction against the developer's need for flexibility.