September 1, 2026
On August 31, 2026, GitHub published its Copilot changelog covering the August releases for VS Code, versions 1.132 through 1.135. It's a dense list. There's a side-by-side chat arrangement so you can view multiple concurrent chats in horizontal or vertical groups, with the layout saved between sessions. There's a `/btw` command for opening a side conversation that shares context and cache with your main chat. There's prompt timeline navigation, so you can jump back to a specific prompt and see the file changes tied to it. There's a portable "Agent Plugins 1.0" standard so customizations work across compatible clients. There's an Offline Agents window that authenticates with an API key instead of a GitHub login. You can switch model providers — Anthropic or a Copilot subscription model — in the middle of a session. You can pick up a Copilot or Claude session that was started in a different app. A new "Agent Host" links multiple VS Code windows to the same running agent session. There's an experimental `/rubber-duck` command that brings in a second model to review the first model's output. And on top of all that: conversation search with regex support, a markdown diff hybrid editor, and per-model token usage broken down by input, cached, and output.
Look past the feature names and a pattern jumps out: almost none of this is about writing code faster. It's about managing the aftermath of code an agent already wrote. Multiple concurrent chat sessions need a layout system because engineers are now running several agents at once. A prompt timeline exists because someone has to be able to retrace which prompt produced which change, days later, across a growing diff. The Agent Host links windows together because a single agent session has become something a team needs to hand off, resume, and share — not something one person just runs and forgets. And `/rubber-duck` is the clearest tell of all: it's a command whose entire job is getting a second AI model to check the first one's work, because nobody fully trusts a single model's output on its own.
None of this is being built for someone who wants a working internal tool. It's built for a professional developer who has to track, review, and reconcile AI-generated source code across sessions, across windows, and increasingly across different model providers. That's a real and specific job, and GitHub is building real infrastructure for it.
We're not going to pretend this release is anything other than sophisticated. Session-linking across windows, regex-searchable conversation history, live token accounting per model — that's serious engineering, and for a team of developers actually shipping AI-assisted code at scale, it's a legitimate step forward. If your job is reviewing diffs an agent generated three sessions ago, you want a prompt timeline. If you're running Claude in one window and a Copilot subscription model in another, you want them to talk to the same session state. This isn't tooling built for a problem that doesn't exist — it's tooling built for a problem that's getting bigger as more code gets agent-generated in the first place.
Our point is narrower than "this is overbuilt." It's that the entire category of problems this release solves — stale context across sessions, drift between what an agent did and what a human remembers approving, disagreement between two models on the same output — only exists because there's generated source code sitting between the request and the result. Every feature on this list is downstream of that fact. Take away the generated code, and most of this infrastructure has nothing to manage.
This is where ViibeStack's AI app builder works fundamentally differently, and it's worth being precise about why. When you describe what you need in ViibeStack, you're not generating source code that then needs a prompt timeline, a diff viewer, and a second model to sanity-check it. You get fields, tables, and workflow steps that are visible and editable directly — the same object a person can look at is the object that runs. There's no separate layer of generated code sitting behind the interface that only an engineer can safely interpret, so there's nothing to reconcile across sessions and no diff to reconstruct context around. That's not a smaller version of what Copilot's session tooling does — it's a different category of build entirely, which is the whole argument laid out in Buy vs. Build vs. ViibeStack.
That distinction matters most for the audience actually reading a Copilot changelog and wondering if it applies to them. If you're a business owner or ops lead who needs a working internal tool — a CRM view, an approval workflow, a booking system — you were never the audience for prompt timelines or model-provider switching. Those features solve a developer's problem: verifying that generated code did what it was supposed to do. You have a different problem: getting a specific tool built and working, with logic you can look at and change yourself, as covered in how ViibeStack works.
We're not arguing GitHub Copilot's approach is wrong, or that this release is overkill. For teams of engineers who are already committed to writing and reviewing AI-generated code at real scale, this kind of session infrastructure is a genuine and growing need, and this month's release is a legitimate improvement for that work. The honest conclusion is narrower: this release is unambiguously built for developers managing code, and that's a different problem from a business needing a specific tool that works. If you're solving the first problem, tools like this are worth learning. If you're solving the second, all of this machinery — the timelines, the rubber-duck reviews, the multi-window session linking — is overhead you can skip entirely, which is really the case we made in The AI Agent Permissions Gap: the more a tool generates on your behalf, the more process you need just to trust it.
Sources