Foundry Toolkit for VS Code at //build: Hosted Agents End-to-End, a Smarter Toolbox, and More

We’re thrilled to announce the latest updates to the Foundry Toolkit for Visual Studio Code unveiled at //build 2026. Since our launch, the toolkit has been constantly evolving, and this release marks a significant milestone. The standout feature? A fully integrated Hosted Agent experience, allowing you to scaffold, run, deploy, and monitor everything from within VS Code. Additionally, we’ve broadened the Toolbox with native enterprise integrations and introduced a range of LangGraph samples, offering every developer a straightforward route from concept to production. Whether you’re just starting or refining a production-ready, observable agent, the Foundry Toolkit is here to support you.

While building an agent is straightforward, transitioning it from a rough idea to a reliable, observable service is crucial. This release provides the complete Hosted Agent lifecycle in VS Code, aligning perfectly with your workflow — scaffold, run, deploy, observe.

Creating a Hosted Agent now begins with an enhanced scaffolding experience and a diverse selection of samples. You can choose from an array of functional templates rather than starting from scratch. The process is more intuitive too: we automatically select your subscription if there’s only one, clarify tab navigation, and tighten spacing for a more streamlined setup.

New Hosted Agent scaffolding dialog with the rich sample picker open

Run (F5) — Inspect as You Build

Hit F5 and your agent will run locally, complete with the Agent Inspector. This new feature aligns with the rest of the extension and showcases Copilot SDK visualisation, allowing you to see real-time insights as your agent runs. It’s the quickest way to validate changes before anything is deployed.

Deploy — A Fresh User Experience

Every team has its own deployment style, so we’ve revamped the user experience and introduced two new deployment options for Hosted Agents:

• ZIP Code Deploy: Package your source code as a ZIP file and deploy it directly to Microsoft Foundry Agent Service.

• Bring-Your-Own-Image (BYOI): If you already have a pre-built container in your Azure Container Registry, you can deploy directly from there.

Hosted Agent deploy dialog showing “ZIP code deploy” and “Bring-your-own-image (ACR)” side by side.

Observe — Ensure It Works in Production

Once your agent is deployed, you can fully monitor its performance:

• Hosted Agent Tracing: Examine end-to-end traces of Hosted Agent invocations directly within VS Code. This includes tool calls, delegation chains, and timing for accurate debugging, eliminating the guessing game.

• Continuous Evaluation Settings: You can now set up ongoing evaluation for deployed Hosted Agents, ensuring quality is consistently monitored, rather than just at the time of deployment.

• Evaluations Node: Easily access evaluation runs and results right from the Foundry project tree with one click.

Hosted Agent trace view showing a span tree of tool calls and timings.

A Toolbox gives your agent its capabilities — it’s the curated set of tools, knowledge sources, and integrations your agent can call on during runtime. Instead of manually connecting each piece, you create your Toolbox once and your agent can reliably access it during both local runs and in production. The outcome? Agents that can utilise real enterprise data and systems, with all connections managed centrally.

• Create: Kick off a new Toolbox from the “Tools Catalog” in the Foundry Toolkit sidebar and select the capabilities your agent requires.

• Connect: Set up enterprise systems through native, first-class connections once, and use them across all your agents.

• Consume: Reference your Toolbox from your Hosted Agent so that its tools are ready to use as soon as the agent runs, both locally (F5) and once deployed.

Building on this process, the Toolbox is now even more robust and tailored for enterprise applications:

• WorkIQ as a Built-in Tool: Experience a first-rate WorkIQ performance boosted by A2A connections — no fallback required. Creating a toolbox with WorkIQ is seamless from the start.

• Fabric IQ (OneLake Catalog) Integration: Seamlessly connect your agents to Microsoft Fabric OneLake catalogs straight from the Toolbox.

• Toolbox Guardrails: Implement content-safety guardrails in your Toolbox for safer agent execution.

• Faster Discovery: Utilise a new Toolbox Search Toggle and Agent Tool Multi-Select to quickly find and integrate multiple tools in one go.

Redesigned Tools Catalog, including WorkIQ and Fabric IQ tiles.

For LangGraph developers, we’ve added five new Hosted Agent samples that align LangGraph with the Agent Framework Responses learning path. This means you can enjoy a comparable, step-by-step guide regardless of your preferred framework:

• MCP: Tool loading from a remote MCP server (defaulting to GitHub Copilot MCP) via MultiServerMCPClient.

• Workflows: A custom StateGraph chaining together three specialised LLM nodes: slogan writer, legal reviewer, and formatter.

• Files: Incorporate local filesystem tools alongside the Foundry-Toolbox code_interpreter that operates on session-uploaded files.

• Human-in-the-Loop: A StateGraph that drafts a proposal and pauses for user approval via langgraph.types.interrupt.

• Observability: Utilise GenAI OpenTelemetry tracing with enable_auto_tracing(); spans, metrics, and logs flowing directly to Application Insights.

We have also updated the existing bring-your-own LangGraph samples to conform with the new hosting layer. This includes chatting with local tools, Foundry-managed Toolbox loading, and SSE-streamed multi-turn sessions supported by a MemorySaver checkpointer, ensuring that all samples reflect how Hosted Agents operate effectively.

Workflow visualisation of the LangGraph human-in-the-loop sample paused at an approval node.

This release is about more than just new features. It includes a redesigned Prompt Builder “Improve an Instruction” dialog for quicker iterations, enhancements for MCP toolbox tool icons, clearer error messages during ZIP-deploys, and various Agent Builder and Playground regression fixes, making the overall experience smoother from start to finish.

We encourage you to share your projects, report issues, or suggest features on our GitHub repository. We can’t wait to see the innovations you create.

Welcome to the next chapter of AI development!