One million tokens of context: the end of forgetting mid-conversation

The biggest technical innovation in VS Code 1.123 is support for a one-million-token context window (1M tokens) for selected models from Anthropic and OpenAI — specifically Claude Opus 4.7 and GPT-5.5. To put that in perspective: one million tokens equals roughly 750,000 words, which is approximately three thousand standard pages of text or the entire codebase of a medium-sized application.

In practice, this means the model won't "forget" during a long chat session — it will maintain awareness of the entire codebase, all previous instructions, and the context of pull requests and issues. For developers working on large-scale projects, this is a significant practical benefit: no more manually trimming context and re-explaining what has already been said.

There is one caveat, though. Microsoft fairly warns that larger context windows consume more tokens per interaction. For teams with usage-based billing (which GitHub Copilot has been increasingly promoting this year), this means more careful cost management. In real terms, the price per query can easily rise by tens of percent — with daily use, these are amounts that the budget of a smaller Czech startup can definitely feel.

Session Sync and Chronicle: your agentic diary

Probably the most interesting new feature from a team collaboration perspective is session sync. VS Code now automatically backs up all chat sessions to your GitHub account. Each session contains the complete conversation history, a list of files touched, repository context including the branch and timestamps, and links to pull requests, issues, and commits.

Session sync is complemented by Chronicle — an interface accessible via the /chronicle command in chat, which lets you search session history in natural language, generate standup reports, look up past work by topic or file, and receive personalized productivity recommendations.

Mitch Ashley from The Futurum Group aptly commented: "Capturing the conversation, affected files, and repository context behind each agent session transforms the editor into a system of record for how software is created. The development process becomes observable and the reasons for changes become traceable. The origin of a change is just as important as the change itself."

For Czech companies starting to experiment with agentic development, this is key infrastructure: the ability to audit AI agent outputs is a prerequisite for deployment in regulated industries such as banking or healthcare.

Agents Window: compare agent outputs side by side

The standalone Agents Window — a companion interface for launching and controlling agent sessions — now allows you to open multiple sessions simultaneously side by side. Developers can drag sessions from the list, use Alt+click, or use the "Open to the Side" context menu.

Only one session is active at a time — it determines what the Terminal, Files, and Changes panels display. Sessions can be pinned so they don't get overwritten when selecting another. A practical scenario: in one window, you review an agent's finished code, while in another, you give it a new task. The feature is currently in preview, but it addresses a real need — comparing the outputs of two agent runs without constant switching.

Research Agent: deep exploration without code changes

Another new feature in preview mode is the Research Agent. Launched via /research in chat, it conducts deep research on a topic across the codebase, relevant GitHub repositories, and the web. The output is a thorough Markdown report with citations.

The agent is designed for depth, not speed — and importantly, it has read-only access. It explores but does not modify code. It is suited for cases where a quick chat answer isn't enough: understanding an unfamiliar library, comparing implementation approaches, or analyzing the behavior of a specific API. For now, the Research Agent is available only in Copilot CLI sessions in the Insiders version, so we'll have to wait for production deployment.

Integrated browser: screenshots and favorites

VS Code's built-in browser received two small but practical improvements. The address bar now lets you save pages to favorites and quickly access them from a redesigned interface that also shows open tabs.

In the screenshots area, two new options have been added alongside the existing "Add Screenshot to Chat": capture selected area and capture full page (the latter is experimental for now and requires enabling in settings). For frontend developers who frequently debug visual issues, this is a welcome time-saver — instead of switching to an external browser and manually cropping screenshots, just a few clicks directly in the editor do the job.

Safety net: a two-hour update delay

VS Code 1.123 introduces a two-hour delay before automatic installation of extension updates. The goal is to create a buffer against problematic or potentially compromised versions. The two-hour window gives the community time to discover an issue before the update spreads to millions of users.

Important: extensions from Microsoft, GitHub, and OpenAI are exempt from this delay and update immediately. Manual updates can, of course, be triggered at any time — no one is forced to wait if they want the latest version right away.

What it means for Czech developers

Visual Studio Code is the dominant development tool in Czechia — according to the Stack Overflow 2025 survey, it is used by over 75% of professional developers. It is free, runs on Windows, Mac, and Linux, and offers full Czech language localization.

The new features in version 1.123 come at a time when agentic coding is experiencing skyrocketing growth. Tools like GitHub Copilot, Claude Code, and Cursor are competing for developers' favor, and companies like Microsoft, Anthropic, and Google are massively investing in infrastructure for agentic workflows.

For Czech companies considering adopting agentic development, VS Code 1.123 delivers a key building block: traceability and auditability. Thanks to session sync and Chronicle, for every code change you can retrospectively determine which agent proposed it, in what context, and based on which conversation. This is essential not only for security but also for meeting the requirements of the EU AI Act, which emphasizes the transparency of AI systems.

The agentic future starts in the editor

VS Code 1.123 is not a revolutionary update — and that's a good thing. Microsoft has bet on gradual, thoughtful integration of AI into the developer workflow. Every added feature addresses a specific pain point: one million tokens removes context limitations, session sync solves auditability, Agents Window enables parallel work with agents.

The direction is clear. Copilot is gradually transforming into more of an orchestrator of AI agents than a simple code autocomplete. VS Code is evolving from a text editor into a platform for agentic development — and version 1.123 is one of the most significant steps in that direction in the past year.