GPT-6 Astra now supports asynchronous tool calls and mid-turn guidance, allowing Agents to receive new requirements while a task is running. But we have noticed that the experience depends not on whether the model can receive a message, but on whether the system can determine whether to cancel, queue, or parallelize the work—and reliably manage tasks that are still unfinished.
What this is
Asynchronous Agents—intelligent agents that can execute tasks while receiving new user instructions—are moving from concepts to product capabilities. Astra can start a long-running query first, then handle work that does not depend on its results. Users can also add constraints such as “reduce the budget” while generation is underway.
The real change is happening at the execution layer. When “stop,” a new message, a tool result, and a scheduled task arrive simultaneously, the system needs to route them to different processing paths. An urgent instruction should cancel the current step and quickly establish safe boundaries; supplementary information is better queued and merged; independent, low-cost queries can run in parallel.
The event router can be a lightweight model that first assesses each event’s priority and relevance. Task cancellation still has limits, however: receiving a “stop” message does not mean that external actions already initiated will be automatically undone.
Industry view
We believe asynchronous capabilities will significantly improve how Agents feel to use, especially for long-running tasks, real-time customer service, and multi-tool collaboration. They also clarify the division of responsibility between model APIs—the interfaces developers use to invoke model capabilities—and runtimes: the model is responsible for understanding changes, while the runtime handles scheduling, state, and result attribution.
The risks are equally clear. First, native asynchronous support raises the bar for model training, tool naming, and system prompts; the capability does not automatically become reliability. Second, unfinished tools must retain an explicit status, rather than presenting “started” as “completed.” Third, batched events can cause models to miss earlier requirements, such as responding only to the final message.
Evaluation therefore cannot focus solely on whether an answer is correct. It must also test whether the system correctly classifies events, cancels dangerous actions, and fully executes when multiple constraints arrive at once. Competitive advantage will gradually shift from the model itself toward task orchestration and state management.
Impact on regular people
For enterprise IT: As enterprises deploy Agents, the focus will shift from connecting models to building task queues, access controls, and audit mechanisms. Only systems that can reliably handle “the requirements changed halfway through execution” are suitable for core business processes.
For individual work: Future office assistants may let us modify budgets, headcount, or output formats midway through a task, without repeatedly canceling and restarting it. But whether “change the plan” can be separated from “stop execution” will still depend on the specific product.
For the consumer market: Voice assistants and intelligent customer service systems will reduce waiting and the need to repeat information. However, actions involving orders, payments, or sending content must not be executed automatically based only on a casual “handle that while you’re at it.” Cancellation and confirmation mechanisms will directly affect user trust.