In this lesson

Five changes from Microsoft's release communications feed — some rolling out now, others arriving in June and July 2026. Each one is summarised, with a plain-English takeaway and a question to consider for your organisation.

Learning objectives
  • Understand how Copilot Chat history is being scoped by experience endpoint (GA June 2026)
  • Describe the three Teams AI Interpreter Simultaneous mode enhancements (GA July 2026)
  • Explain how Teams Facilitator detects and answers unanswered meeting questions (GA July 2026)
  • Assess the practical impact of the Copilot Calendar Agent for M365 Copilot users
  • Communicate the OneDrive recycle bin behavioural change to your team accurately
Items 01–05 · Microsoft 365 Changes
01 · M365 Change
Copilot Chat History Filtered by Experience Endpoint
Chat history in Microsoft 365 Copilot Chat will be scoped to the specific Copilot experience the user is currently in — Teams, browser, or Office apps. Users will see only conversations from that context by default, with an option to switch to an All chats view. General availability is June 2026.
Users who move between different Copilot surfaces will no longer see all their chat history by default. They need to switch to All chats to find conversations started in a different context.
Does your team use Copilot Chat from multiple entry points? Do they know where to find history across those contexts?
Source: Microsoft 365 Roadmap #559601
02 · M365 Change
Teams AI Interpreter Gets Caption Sync and Admin Voice Controls
Three enhancements arrive for AI Interpreter's Simultaneous mode. Audio and live captions will now use the same language, ending the previous mismatch. Admins will be able to disable voice simulation entirely. Users get more voice options with dynamic speaker assignment. General availability is July 2026.
Multilingual meetings will be easier to follow with consistent audio and caption language, and IT admins gain explicit control over whether voice simulation is enabled.
Does your organisation use Teams AI Interpreter? Have your admins reviewed the new voice simulation disable option?
Source: Microsoft 365 Roadmap #562035
03 · M365 Change
Teams Facilitator Detects Unanswered Meeting Questions
Teams Facilitator will automatically detect when a question raised during a meeting goes unanswered. It posts a prompt offering to find an answer using web search. Participants choose whether to accept. General availability is July 2026.
The Facilitator acts as a quiet backstop for factual queries in meetings, reducing post-meeting follow-ups on basic informational questions.
In your regular meetings, how many questions typically go unanswered or require research after the call ends?
Source: Microsoft 365 Roadmap #558341
04 · M365 Change
Copilot Calendar Agent Manages Meeting Rules in Plain English
The Copilot Calendar Agent is rolling out for M365 Copilot licence holders. Users set plain-English rules — for example, decline meetings with less than 24 hours' notice — and Copilot acts on them automatically in the background. Every action is reviewable and adjustable at any time.
For professionals managing high meeting volumes, this is a practical way to enforce scheduling preferences without manually handling each invitation.
If you could set three plain-English rules for your calendar today, what would they be?
Source: eMDTec / Microsoft 365 May 2026
05 · M365 Change
OneDrive Remote Deletes No Longer Appear in Local Recycle Bin
From early May 2026, files deleted from OneDrive via the web or a different device will not appear in the local Windows Recycle Bin. They remain recoverable via the OneDrive web recycle bin. Files deleted locally on the same machine are unaffected. There is no in-product notification of this change.
Anyone checking their local Recycle Bin after a remote delete will not find the file there. Recovery now requires the OneDrive web interface.
Have your users been told about this change? Is there a support doc they can reference if a file appears to go missing?
Source: eMDTec / Microsoft 365 May 2026
Quick review — accordion
01 · Copilot Chat History Filtering ▼
Chat history scoped to current Copilot experience endpoint. Default view shows only conversations from that surface. "All chats" view available. GA June 2026. Ask: Does your team know where to find cross-surface history?
02 · Teams Interpreter Enhancements ▼
Caption and audio language now match. Admins can disable voice simulation. More voice options with dynamic speaker assignment. GA July 2026. Ask: Has your IT admin reviewed the disable option?
03 · Facilitator Answers Questions ▼
Detects unanswered questions in Teams meetings. Offers web search answer. Participant accepts or ignores. GA July 2026. Ask: How many meeting questions go unanswered in your team?
04 · Copilot Calendar Agent ▼
Plain-English scheduling rules acted on automatically. Available to M365 Copilot licence holders. All actions reviewable. Rolling out now. Ask: What three rules would you set?
05 · OneDrive Recycle Bin Change ▼
Remote/web deletes no longer appear in local Windows Recycle Bin. Recover via OneDrive web. Local deletes unchanged. No in-product notice. Ask: Have you communicated this to your users?
Knowledge check
A user deleted a file from OneDrive via the web browser on their laptop. A colleague on a desktop PC checks their local Windows Recycle Bin for the file. What will they find?
The file will be in the local Recycle Bin as normal
The file cannot be recovered at all
The file will not be in the local Recycle Bin — it must be recovered via the OneDrive web recycle bin
The file appears in both the local and web Recycle Bin
Full citations: see 260523-D3F4-sources.json in the same folder. Primary source for items 01–03: Microsoft 365 Roadmap. Items 04–05: eMDTec verified summary of Microsoft release notes, May 2026.
1 / 4
In this lesson

Three developments that matter for anyone thinking about how AI agents are moving from experimental to production environments. Items 06–08 cover Anthropic, Camunda, and Google — three different angles on the same underlying shift.

Items 06–08 · AI & Productivity
06 · AI & Productivity
Anthropic Adds Self-Hosted Sandboxes and MCP Tunnels to Claude Agents
Anthropic released public-beta self-hosted sandboxes for Claude Managed Agents on 19 May 2026. Enterprises can run agent tool execution on their own compute via Cloudflare, Daytona, Modal, or Vercel. A research-preview MCP tunnels feature lets agents call internal MCP servers via an outbound-only encrypted gateway, keeping sensitive systems behind the firewall.
Regulated organisations that could not use managed agents due to data residency or security perimeter requirements now have a more viable path to production deployment.
Does your current security policy prevent AI agent adoption? Would keeping tool execution inside your own infrastructure change that?
Source: AI Agent Store — Anthropic update, 19 May 2026
07 · AI & Productivity
Camunda Launches ProcessOS for Governed Agentic Workflows
Camunda announced ProcessOS at CamundaCon, now in closed beta from 20 May 2026. It discovers existing business processes, re-engineers them as agentic workflows, and optimises them continuously. Built-in human review gates are included, with AWS and Bedrock integrations.
Operations teams managing complex ERP or CRM stacks can describe an outcome and have ProcessOS convert it into a repeatable, auditable agentic process without rewriting existing integrations.
Which single high-value, lower-risk process in your organisation would you want to convert to a governed agentic workflow first?
Source: AI Agent Store — Camunda CamundaCon, 20 May 2026
08 · AI & Productivity
Google Releases Gemini 3.5 Flash and Antigravity Agent Platform
At Google I/O on 21 May 2026, Google announced Gemini 3.5 Flash — lighter and faster, tuned for agentic workflows — alongside the Antigravity developer platform. Antigravity lets developers build persistent background agents across Workspace, Search, and developer tools at lower cost than GPU-only stacks.
For teams evaluating multi-provider AI strategies, Antigravity provides an agent development surface using existing Workspace integrations as the natural entry point.
If persistent background agents were available in your tools today, which daily routine would you hand off first?
Source: AI Agent Store — Google I/O 2026, 21 May 2026
Three angles — tabs

Anthropic: Self-hosted sandboxes let enterprise tool execution stay inside the customer's own compute. MCP tunnels create an encrypted outbound-only channel for calling internal tools without opening inbound firewall rules.

Camunda: ProcessOS is an orchestration layer that sits above existing ERP/CRM systems, discovers processes from observation, and converts them into governed agentic workflows with audit trails.

Google: Gemini 3.5 Flash is a smaller, cheaper model for multi-step agentic tasks. Antigravity is the developer platform for building persistent agents that run in the background across Workspace products.

All three announcements address the same underlying obstacle: enterprises want to deploy AI agents in production, but face blockers around security, governance, and integration cost. Anthropic lowers the security barrier. Camunda lowers the integration barrier. Google lowers the cost barrier.

The common direction is agents that operate inside existing systems rather than alongside them. Each announcement reduces a reason to wait.

For Anthropic: If security is the blocker, request access to the MCP tunnels preview. Run one low-risk, read-only agent to validate audit logs and incident response before expanding scope.

For Camunda: Register for the ProcessOS closed beta. Pick one high-value, low-risk process — something like routing or triage — to pilot with human review gates in place.

For Google: If your team already uses Workspace, trial Antigravity on a single contained workflow (calendar plus follow-up email) to benchmark cost, latency, and where approvals are needed.

Key terms — click to flip
Term
MCP Tunnels
Anthropic feature that lets agents call internal MCP servers via an outbound-only encrypted gateway, without opening inbound firewall rules.
Click to flip back
Term
Self-hosted Sandbox
Execution environment that runs agent tool calls on customer-managed or partner compute (e.g. Cloudflare, Vercel), keeping data inside the customer's security perimeter.
Click to flip back
Term
ProcessOS
Camunda's AI orchestration layer that discovers existing business processes and re-engineers them as governed agentic workflows with built-in human review gates.
Click to flip back
Term
Antigravity
Google's agent-first developer platform, announced at I/O 2026. Supports persistent background agents that run across Workspace, Search, and developer tools.
Click to flip back
Knowledge check
Which of the following best describes why Anthropic's MCP tunnels feature matters for regulated enterprises?
It reduces the cost of running Claude models by routing traffic through cheaper infrastructure
It lets agents call internal tools without opening inbound firewall ports, keeping sensitive systems inside the security perimeter
It allows Claude to access public internet APIs without any authentication
It enables Claude to run entirely offline without any cloud connection
Sources for items 06–08: AI Agent Store daily briefing, 19–21 May 2026 (aiagentstore.ai). Original announcements from Anthropic, Camunda, and Google respectively. Full citations: 260523-D3F4-sources.json.
2 / 4
Item 09 · Editorial Insight
09 · Insight
The Agent Is Now Inside the Tool
This week, AI agents embedded inside existing tools rather than asking people to open separate apps. Teams Facilitator answers questions inside the meeting. Copilot Calendar Agent acts on rules without being invoked. Anthropic MCP tunnels reach internal systems. Camunda ProcessOS watches and restructures workflows. Adoption is becoming passive — agents act without being called. That raises a question most organisations are not structured to answer: who reviews what the agent decided, and when? The audit trail question is not technical. It is organisational. Someone needs to own the log of agent-initiated actions, and most teams do not have that role defined yet.
Before enabling any of these agent features, decide who owns the review of agent-initiated actions. Every automated decision is a decision — and without a defined owner, errors compound quietly.
Does your organisation have a process for reviewing actions taken by AI agents on behalf of users? If not, who should own that?
Key takeaway
Before enabling agent features, decide who owns the review of agent-initiated actions. Without a defined owner, errors compound quietly.
Three steps to act on this
01
Notice — what pattern do you see?
Look at the five M365 items and three AI items. How many involve an agent or automated process acting without explicit invocation? Which of these features does your organisation currently use or plan to enable?
02
Experiment — one thing to try this week
If you have access to Copilot Calendar Agent, set one rule and monitor what actions it takes over 48 hours. Review the action log. Is there anyone in your team who would have made a different decision? If so, write down what the conflict was.
03
Adapt — what would you change in how you work?
Draft a one-paragraph policy statement for agent-initiated actions in your team. Who receives the log? How often is it reviewed? What is the escalation path if an agent decision causes a problem? Even a rough draft is better than none.
Editorial note

This is the editorial view. It connects observable patterns from the sourced items but represents a judgement call, not a fact. Check the sources and form your own assessment.

3 / 4
What we covered in this edition

Five changes from the Microsoft 365 release feed — Copilot Chat history scoping, Teams Interpreter enhancements, Teams Facilitator question answering, the Copilot Calendar Agent, and the OneDrive recycle bin change. Three AI developments — Anthropic self-hosted sandboxes and MCP tunnels, Camunda ProcessOS, and Google Gemini 3.5 Flash with the Antigravity platform. One editorial insight connecting the pattern: agents are embedding inside tools, adoption is becoming passive, and most organisations lack a defined owner for agent-initiated action review.

Key terms — click to flip
Term
Copilot Calendar Agent
M365 Copilot feature that applies plain-English scheduling rules automatically in the background. Requires M365 Copilot licence. All actions are reviewable.
Term
Teams Facilitator
Teams agent that automatically detects unanswered questions during meetings and offers to search the web for an answer. GA July 2026.
Term
Simultaneous Interpreter Mode
Teams AI Interpreter feature providing real-time multilingual audio and captions. Updated May 2026: captions now match audio language; admins can disable voice simulation.
Term
Copilot Chat History Scoping
June 2026 change that limits the default chat history view to conversations from the current Copilot experience endpoint. "All chats" view allows cross-surface history.
Term
Gemini 3.5 Flash
Google model announced at I/O 2026. Lighter and faster than previous Gemini versions, tuned for multi-step agentic workflows with lower cost than GPU-only stacks.
Term
Passive Adoption
The editorial insight's term for AI agent behaviour that acts without explicit user invocation — creating governance gaps when no one owns the review of agent-initiated decisions.
Final knowledge check
According to the editorial insight in this edition, what is the most important organisational step before enabling AI agent features that act on behalf of users?
Get approval from the IT security team to whitelist the AI vendor's domains
Ensure all users have the correct Microsoft 365 Copilot licence tier
Define who owns the review of agent-initiated actions and how the audit trail is managed
Wait for all features to reach general availability before enabling any of them
Closing
Nine items. Sourced, summarised, and delivered. The next edition runs at 18:00 today. Same run, same format, updated feed.
Run ID: 260523-D3F4 · Generated: 23 May 2026 AM · Sources: Microsoft M365 RSS + verified web · Full citations: 260523-D3F4-sources.json · Catalogue: ../CATALOGUE.html
4 / 4