If two independent tasks wait for each other, I want a better explanation than a habit of running everything in sequence. GPT-5.6's general availability announcement on July 9, 2026 included an ultra mode aimed at parallel agent coordination. That promise matches a very concrete expectation I have about getting work done.
Two days later, on July 11, 2026, I asked why the agents were not doing things in parallel. I expected tasks to move at the same time. My question concerned progress; the ultra mode announcement offers an implementation of that idea which still needs to be judged by what it can deliver in practice.
Requesting parallelism takes one sentence. Dividing work properly requires understanding the problem. Two parts may depend on the same unresolved decision. Assigning an agent to each side without recognizing that dependency can produce work to undo. I want coordination to find what can already proceed and make clear what has to wait.
In a hypothetical example, investigating two implementation alternatives lets each agent read different material and return a comparable proposal. There is a reason to work simultaneously. If both need to edit the same section before the choice is made, the division requires isolation and responsibility for changes. Opening another conversation does not automatically resolve that conflict.
A panel can be full of active agents while the task still lacks an acceptable delivery. I want to know what each execution resolves and how the answers will be assembled. If I have to reconcile the conclusions manually, the benefit remains incomplete. Integration belongs in the task's duration, even when individual investigations finished early.
Waiting periods deserve attention too. While one answer is pending, there may be independent work to advance. I want the agent to look for that opportunity. When none exists, an explanation of the dependency is enough. That helps me distinguish a necessary wait from a sequence chosen because nobody considered another way to distribute the request.
I will evaluate parallel coordination with a task whose parts are clear and follow it through to an assembled delivery. The calculation includes the review needed to accept the whole result. My July expectation still applies: show which parts can advance together and return their results in a form I can check. If ultra mode helps with that, it has a concrete use for me. The number of open agents alone does not change the task's status.