Parallel agents in ChatGPT: how to organize a first test at your company

Delegating a task to AI requires explaining the expected result. When different agents work on parts of that task at the same time, another responsibility comes up: defining how the deliverables will fit together.

On October 16, 2026, OpenAI logged a GPT-6 Pro update in ChatGPT with parallel execution by multiple agents, progress tracking and access to completed tasks. The announcement also reports faster answers and improvements in the use of plugins and tools. Source: OpenAI.

In DG5i's assessment, the first test should involve internal work, with a verifiable result and no automatic external actions. Preparing a comparison of sales proposals is one possible example.

Imagine a company that received documents from different vendors. One track could organize the items included in each proposal. Another could gather deadlines and support terms. Both would work on the same authorized set of files, but with separate deliverables.

Then a person would bring the results together, check the cited passages and identify missing information. The choice of vendor would stay with whoever knows the company's priorities.

This scenario is hypothetical. It is not a test that was carried out, nor a guarantee that any file, integration or flow will be available in the user's account.

Asking only to “compare the proposals” leaves important decisions open. A more precise request establishes what to extract, where to look and how to handle gaps.

In the example, each deliverable could have four fields: vendor, information found, source document and open item. If a deadline isn't given, the result should say “not provided”, without filling the gap by assumption.

Using a common format makes consolidation easier. It also helps notice when two tracks interpreted the same condition in different ways.

Activities can happen together when none needs the still-incomplete result of another. The final recommendation, however, should wait for the relevant information to be checked.

If a proposal requires clarifying a support condition, that open item needs to appear in the comparison. A missing answer can't turn into an invented advantage or disadvantage just to close the analysis.

That's the point where the manager should step in: decide whether there is enough information to move on or whether the work needs to stop until someone clears up the question.

For a first test, the recommendation is to use only the necessary and authorized documents. The request should exclude sending messages, changing systems, hiring or any commitment on the company's behalf.

Tracking can record the total time until review, the number of corrections and the open items the agents made explicit. Compare these records with the usual process before concluding there was a gain.

A fast run that requires redoing the analysis doesn't demonstrate savings. A deliverable that lets you check each piece of information, on the other hand, can be useful even when it needs a few adjustments.

The release covers eligible Pro users in the Chat tab and depends on workspace rules in managed environments. This doesn't prove access in the company's specific account or fit with its data and processes. Source: OpenAI.

The best first step is a small test: a bounded task, authorized sources, independent parts and human review before any external decision.

From that experience, the company can decide where parallel work helps and where the traditional sequence is still necessary. The choice should come from the quality of the reviewed deliverable, not just from the speed shown during execution.