Foundry Toolkit for VS Code — August 2026 Update

Welcome to the August summary of the Foundry Toolkit for VS Code. This month saw the launch of four updates: 1.6.7, 1.6.8, 1.6.9, and 1.6.10.

In August, our focus was on creating a seamless workflow for agent development. We aimed to provide a clear path from start to finish, enabling you to connect reusable tools, run with genuine user isolation, and analyse exactly how time and tokens were utilised.

If you have any feedback or encounter a bug, please report an issue on GitHub. Your insights help shape our roadmap!

• Prompt Agent Toolboxes — You can now connect to a centrally managed toolbox, review its tools and skills, manage versions, and adjust approval policies right from Agent Builder. 1.6.10
• Agent-to-Agent Connections (Preview) — Easily connect any A2A-compatible agent using a configured connection, the Foundry account catalog, or a custom HTTPS endpoint. 1.6.10
• Agent Inspector Overview — Access a visualisation of latency and an ordered timeline showing model activity, reasoning, and tool use for each run or for a selected run. 1.6.9
• User-Scoped Hosted Agent Sessions — Set individual user identities to keep Responses conversations and session files separate for every user. 1.6.8
• Improved Create Agent Start Page — Now you can easily select from Microsoft Agent Framework, Copilot SDK, LangGraph, or Codex through a redesigned interface. 1.6.9

Starting your agent development journey shouldn’t mean making the wrong choices right off the bat. The newly redesigned Create Agent page directs you straight to options like Microsoft Agent Framework, Copilot SDK, LangGraph samples, Copilot-assisted coding, and Agent Builder. You have the flexibility to choose whether you want to dive straight into coding, follow a guided build, or start with a prompt agent — all before getting stuck in the details. 1.6.9

Setting up hosted agents is now more straightforward. You can select Skip for now if model setups fail to load, allowing you to connect the model later. Also, administrator-connected Foundry models now show up alongside standard deployments in playgrounds and during Hosted Agent creation, ensuring you can access existing configurations easily. 1.6.8 1.6.9

Once your agent is up and running, user identity becomes crucial. The Hosted Agent Playground now lets you assign a unique user identity for Responses conversations. This keeps conversation states and session files distinct for each user instead of merging everything into one session. Additionally, if someone sends you a Microsoft Foundry portal link, the deep links will direct you straight to that Hosted Agent’s Details or Optimisation page in VS Code — avoiding the hassle of navigating from the portal home or search screen. 1.6.8 1.6.10

Managing an agent with multiple tools can often lead to redundant configurations and approval processes. However, the new Toolbox feature addresses that by packaging together centrally managed tools via a single Model Context Protocol (MCP)-compatible endpoint, allowing for shared versioning and policy controls.

In August, we introduced toolbox workflows within Agent Builder. You can now access toolboxes directly by clicking on Add Tools or using the Add to Prompt Agent option from the toolbox resource list. Once you attach a toolbox, it shows as a collapsible card where you can review tools and skills, switch versions, configure approval policies, manage nested tools, or even replace/remove the toolbox as needed. This way, you maintain a well-managed collection instead of a mix of one-off connections. 1.6.10

Agent-to-agent connections have also been enhanced. You can now easily integrate an A2A-compatible agent either from an established connection, the Foundry account catalog, or a custom HTTPS endpoint. Just connect it to a Prompt Agent or place it inside a toolbox for future use across different agents and runtimes. Your development pipeline can now consist of agent → toolbox → specialist agent, which allows for real resource management instead of hiding connections in prompt text. 1.6.10

A final output can sometimes appear correct, yet the process behind it might be inefficient or incorrect. With the enhanced Agent Inspector, you can now access the sequence of actions along with the evidence.

The newly added default Overview tab displays each run or a selected run with two synchronised views: a latency waterfall and an ordered timeline demonstrating model activity, reasoning, and tool usage. Response footers now include model information, total duration, the number of tokens used, and timestamps; you can even hover over the token total to separate input from output. Additionally, any raw reasoning provided by the agent appears in separate collapsible sections. 1.6.9

With the updates in 1.6.10, the tool inspection feature has become even more insightful. Calls are organised by response run, complete with statuses, call IDs, arguments, and results. Now, individual Responses events also indicate when they reached the Agent Inspector. The Overview waterfall and timeline scroll independently, making it easier to track long streaming responses. You can quickly pinpoint issues without needing to piece together the run from chat history. 1.6.10

We’ve simplified conversational interactions as well: use the Up or Down keys to recall and modify previous requests without losing your draft, or select Clear Chat to reset the conversation and its state. Pending MCP approvals and OAuth requests will be displayed at the top of the input area, along with options for bulk actions and expandable details until all decisions are made. 1.6.7 1.6.9

We’ve ensured that resource pages remember your ongoing work rather than resetting it. Models and Tools will now load your previously selected tab first, displaying core details before fetching additional information. This means that when you return to Agents, Models, Tools, Knowledge, or Evaluations, the toolkit retains your rows, searches, filters, and pagination, refreshing the current view in the background. You can still perform a manual refresh if you need the latest service status. 1.6.7

Even the sidebar has been streamlined. The collapsed My Resources sections now load on-demand, while Search and Recent Agents remain easily accessible. Evaluations, Routines, Tools, Skills, and Toolboxes now offer consistent loading feedback, and direct links to Tools or Skills open the appropriate tab without having to load Toolboxes first. This improves efficiency without creating unnecessary waits. 1.6.7 1.6.8

We’ve also fixed a sharp issue with model deployment guidance: quota errors now direct you to the token quota page for your current Foundry project, so you can easily identify your project’s needs. 1.6.10

For those developing Activity Protocol agents for Microsoft 365, the recent updates ensure that local debugging uses the same terminology. Newly created Python projects will now open the Microsoft 365 Agents Playground inside VS Code, allowing you to testactivity-shaped conversations right from the editor before deployment, instead of relying on an incompatible playground. 1.6.7

Copilot-assisted projects are now streamlined too, following the same path as current Hosted Agent setups. This includes the Foundry project and model setup, a managed Python environment, debugging at the workspace-root level, and the latest flow for local runs and deployments. When you reuse an existing Foundry project, Copilot will no longer prompt you to select its Azure location again. 1.6.10

• Agent Inspector — Streamlined response and reasoning text remains intact; the sequence of response text, reasoning, tool calls, and permission decisions are preserved; replacements will phase out old stream events; and unmatched tool calls or results will no longer be visible in Details. 1.6.9 1.6.10
• Approvals and Consent Processes — Human-in-the-loop delays no longer result in duplicate tool or approval cards, Clear Chat is still accessible during waiting periods, and continuation responses will retain pending approvals until every request is addressed. 1.6.8 1.6.9
• Activity Protocol Deployment — Azure Bot settings are now validated before submission, compatible Bots are reused effectively, and any identity or application ID conflicts come with recovery suggestions. Successful deployments will no longer open an unsupported Agent Playground. 1.6.7 1.6.8
• Agent Builder and MCP OAuth — Reopening a Foundry Prompt Agent maintains its selected version and tool configuration, while authorization callbacks will only complete the matching connection requests. 1.6.8
• Improved Accessibility — Screen readers now announce actions taken in the Model Catalog, as well as collapsible controls in Agent Builder and Model Preference sections, ensuring that project and model fields display their labels and states clearly; prompt placeholders also meet minimum contrast standards. 1.6.10

• Removal of GitHub Models — The GitHub Models option has been completely removed from the Model Catalog, playground, model comparison, Agent Builder, and evaluations after the service’s retirement. If you attempt to run a saved workflow or evaluation that refers to GitHub Models, simply open it up and select a different model from those available. 1.6.7

This August, we established a comprehensive connection throughout the agent workflow: from selecting the right starting point and reusing structured tools to composing agents through A2A, isolating real user sessions, and meticulously inspecting every aspect of the run, including timing, tokens, arguments, and results.

Give the toolbox a shot with your next Prompt Agent, view the run in the Agent Inspector, and let us know if there are any hiccups in the workflow that slow you down.

Happy building!