While consumer attention this month has been focused on Windows 11’s recovery overhaul, Microsoft’s developer and productivity tooling quietly shipped one of its busier release cycles of the year. Visual Studio Code jumped to version 1.131 with a substantial set of agentic-workflow improvements, Copilot in Excel picked up new grounding and citation capabilities alongside two new frontier model options, and the underlying Agent Framework received a breaking-change overhaul that developers building on top of it will need to plan around. Taken together, the releases paint a clear picture of where Microsoft is pushing its tools: deeper visibility into AI agent behavior, and tighter integration between AI models and the documents people actually work in.
VS Code 1.131: making subagents visible
Visual Studio Code’s agentic features have matured quickly over the past year, but one persistent complaint from developers has been the “black box” feeling of watching a coding agent work without knowing what it’s actually doing at any given moment. Version 1.131 addresses this directly with a new subagent visibility panel. Developers can now see, at a glance, which model a running subagent is using, how long it has been executing, and whether it currently has an active tool call in flight — all without needing to open the subagent’s full conversation transcript.
This might sound like a minor UI tweak, but for teams running multiple agents in parallel across a large codebase, it changes the day-to-day experience considerably. Instead of babysitting a single conversation window, developers can monitor several concurrent agent runs the way they might monitor a CI pipeline dashboard — scanning for the one that’s stuck or taking unusually long, rather than reading through every line of reasoning.
Built-in dictation, no extension required
Also new in 1.131 is experimental built-in dictation support across chat, the code editor, and the integrated terminal. Previously, voice input in VS Code required installing a separate Speech extension and configuring a provider; this release folds basic dictation directly into the editor shell. It’s marked experimental, so expect rough edges, but the direction is notable: Microsoft is treating voice as a first-class input method for prompting agents and writing comments, not just a niche accessibility add-on.
A hybrid Markdown editor for agent collaboration
The other headline addition is a new hybrid Markdown editor, currently scoped to the Agents window. It allows developers to view and edit Markdown files while also attaching agent-actionable comments directly inline — essentially a lightweight review layer that lets a human leave a note in a plan or spec document and have an agent pick it up as an actionable instruction. This fits into a broader pattern across the update: agents are increasingly expected to work from structured planning documents rather than purely ad hoc chat instructions, and tooling is catching up to make that workflow feel native rather than bolted on.
Rounding out the release are a series of accessibility refinements and terminal behavior fixes, continuing the incremental hardening that has characterized VS Code’s monthly cadence for the past several release cycles.
Copilot in Excel: from formulas to full analysis workflows
On the productivity side, this month’s Copilot in Excel update pushes the assistant further away from being a formula helper and closer to being a genuine analysis partner. The most significant addition is inline citations — when Copilot pulls a number or claim into a worksheet, it now shows its sourcing directly in context, rather than leaving users to simply trust an unattributed output. For any organization using Copilot-generated analysis in reporting that gets shared upward or externally, that traceability is a meaningful trust upgrade.
Alongside citations, the update adds:
- Synced connectors — data connections that stay live and refreshed rather than requiring manual re-import, so Copilot’s analysis reflects current data automatically.
- Power BI grounding — Copilot can now reference and reason over existing Power BI models and reports when generating analysis inside Excel, reducing duplication between the two tools.
- Brand and theme design skills — Copilot can apply consistent branding, color themes, and formatting conventions across workbooks automatically, useful for teams producing recurring reports that need to look consistent without manual formatting passes each cycle.
- AutoSave-off support — a small but practical fix: Copilot in Excel now works properly in workbooks where AutoSave has been intentionally disabled, which previously caused friction for users working with sensitive or version-controlled files.
Two new frontier models join Copilot in Excel
Perhaps the most-discussed change is model choice: this release adds OpenAI’s GPT-5.6 and Anthropic’s Claude Opus 5 as selectable models within Copilot in Excel, sitting alongside Microsoft’s existing default options. For analysts working with dense financial models or messy real-world datasets, being able to swap the underlying model for a given task — rather than being locked into a single default — is a meaningful flexibility upgrade, and reflects a broader trend of productivity suites treating model choice as a user-facing setting rather than an implementation detail hidden from the end user.
The Agent Framework’s breaking-change overhaul
Underpinning both of the above is a substantial update to Microsoft’s Agent Framework, spanning both the .NET and Python SDKs. This release includes breaking changes that developers building agentic applications on the framework will need to account for before upgrading production systems. Key changes include:
- New approval and memory capabilities, giving agents a more structured way to request human sign-off before taking consequential actions, and to persist context across sessions.
- Workflow and message-ordering fixes, addressing edge cases where multi-step agent workflows could process messages out of the expected sequence.
- A split of the legacy AGUI package into separate SDK components, intended to make upgrades more modular going forward — rather than pulling in the entire framework for a small dependency change, teams can now update individual pieces independently.
For engineering teams with agents already in production, the practical advice is straightforward: review the framework’s migration notes before bumping the dependency version in any live system, particularly around the AGUI package split, since that restructuring is the change most likely to silently break existing integration code.
Key takeaway: This release cycle is less about flashy new features and more about maturing the plumbing — visibility into agent behavior, traceability of AI-generated analysis, and cleaner SDK boundaries for developers building on top of Microsoft’s agent tooling.
Why model choice inside productivity apps is becoming the norm
The decision to offer multiple frontier models directly inside Excel, rather than locking users into a single proprietary option, reflects a shift that’s playing out across the productivity software industry more broadly. As large language models have become commoditized to a degree — with several vendors now offering models that are highly competitive on any given benchmark in a given month — software vendors have increasingly concluded that betting an entire product experience on one model provider creates unnecessary risk, both technically and competitively. Offering choice lets power users pick the model best suited to a specific task: one model might handle messy, inconsistently formatted financial data more gracefully, while another might produce more concise, presentation-ready summaries. For everyday users who don’t want to think about the distinction at all, a sensible default remains in place, so the added flexibility doesn’t come at the cost of a simpler out-of-box experience for the majority of users who just want Copilot to work without configuration.
This also creates a practical governance question for organizations rolling out Copilot broadly: if employees can select among multiple third-party models for analysis that touches sensitive financial data, IT and compliance teams need clarity on which models are approved for which data classifications, since different providers may have different data-handling and retention terms. Expect Microsoft to publish more explicit admin-level controls over which models are available to which users as this multi-model approach expands to more of the Microsoft 365 suite.
The developer experience implications of agent visibility
Stepping back from the specific feature list, the subagent visibility work in VS Code 1.131 speaks to a maturity inflection point for agentic coding tools generally. Early agentic coding assistants asked developers to essentially trust a single conversation thread and hope for the best; the current generation is increasingly built around the assumption that developers will be running multiple agents concurrently across different parts of a codebase, and that the tooling needs to support genuine oversight at that scale rather than a single-agent mental model. This is a meaningfully different design problem, closer to monitoring a fleet of automated workers than supervising one assistant, and the tooling investments this month reflect that shift explicitly.
What to watch next
With both VS Code and Copilot in Excel converging on the theme of “trust the agent, but verify,” expect future releases to keep expanding observability features rather than raw capability. On the Agent Framework side, the SDK split suggests Microsoft is preparing the platform for a wider range of hosting and deployment scenarios, which is worth monitoring for any team currently building custom agent infrastructure on top of Microsoft’s stack. We’ll keep tracking release notes as they roll out across the broader Microsoft 365 and developer tooling ecosystem.