Technical Context
I wouldn't rush to shout about a "secret system prompt." Based on the data currently available, there is no official confirmation of a hidden update. However, there is a quite down-to-earth version: in May and June, GitHub reshuffled the models in Copilot, and it is precisely after such shake-ups that AI automation starts behaving "as if someone swapped it."
What I see from the facts: GPT-5.3-Codex became the base model for some Copilot scenarios, GPT-5.2-Codex started to be phased out, and the rollout was not instantaneous. It's in these transitions that strange effects arise: yesterday the agent was overly autonomous, today it stalls and interrogates every little detail.
I was particularly intrigued by the piece about memory.md and introspection across threads. This no longer looks like a single button or a single mode. If the assistant has memory, internal instructions, and different product surfaces, then the change in behavior might well come not from a "system prompt" in its pure form, but from a combination of a new model, memory, and orchestration.
And here I would dig not toward theories, but toward reproducibility. On which account did this start, in which mode, on which exact Copilot surface, with which model, what's in memory.md, and does the effect disappear after resetting the context. Without this, all talk about /plan mode remains at the level of speculative chat.
What This Changes for Business and Automation
If you're building AI integration in development, this behavior hits two spots at once: speed and predictability. An agent that previously handled a task in one pass now inserts an extra layer of approvals. On small tasks, this is just annoying. In a team pipeline, it becomes a hidden time tax.
Who might even benefit from this? Teams with strict requirements for quality and safe changes. Those who lose are the ones counting on a fast, semi-autonomous flow of fixes without constant back-and-forth.
I see such shifts regularly: the model is changed, while the AI architecture of the process remains old, and suddenly everything starts stuttering. At Nahornyi AI Lab, we specifically analyze these bottlenecks in practice: where a more autonomous agent is needed, where strict boundaries, and where memory and routing should be completely rethought. If your Copilot or internal code assistant suddenly starts slowing down the team, you can calmly dissect the workflow and assemble AI solution development to fit the actual work, not the expectations from the previous version.