AI & Agentic Weekly Update — October 5, 2026

Autopilot Takes Shape as Scout Steps Aside

Following the September 25 redesign of Copilot, attention this week turned to Autopilot, the persistent agent that replaced the Microsoft Scout name. Autopilot is described as an always-on, proactive agent that holds its own identity, memory, and workspace inside a tenant, and keeps working on an assignment such as a supplier review, scheduling, or stakeholder follow-ups without being re-prompted. It remains in private preview, with wider access expected through the Frontier program in the coming weeks. General availability has not been announced, and several analysts expect more detail at Ignite in November.

The administrative picture matters more than the feature list. Autopilot runs under a governed Microsoft Entra identity with scoped credentials, so admins must decide which resources it can reach, when a human must approve a sensitive action, and how Microsoft Purview policies apply. Email, calendar, and OneDrive collaboration reportedly require a separate optional agent user account, which is currently limited to Frontier preview tenants. Treat each Autopilot as a managed entity with a named owner and an audit trail, in the same way you would treat a new employee account. Also avoid confusing it with Windows Autopilot device provisioning, which is unrelated.

What to do: Before enrolling anyone in the preview, write a one-page standard for Autopilot: who may create one, who owns it, which data it may touch, and which actions require human approval. Name an owner for every instance from day one.

Copilot Cowork: September Releases Reach Admins

The September 2026 update for Copilot Cowork includes several changes worth knowing about. Local browser automation became generally available to all Copilot tenants. A new reasoning control lets users choose how much effort Cowork spends before responding, trading speed for thoroughness. A Cost skill shows month-to-date spending against a budget, and the Scheduled tab was renamed Automations. An App skill that builds interactive apps from a plain description is in Frontier first, and users can refine, publish, and share the results inside the organization.

Frontier models are also now available for delegation-style work in Cowork and Copilot Studio, including GPT-6 Astra and Claude Fable 5.1, with Work IQ grounding responses in organizational context. Cowork is also the long-running half of the new Home destination, where Chat and Cowork merge and an "Auto" routing feature is planned to pick the right capability for a request. Home is expected to begin rolling out in October.

The practical effect is that Cowork is moving from a demo to a daily tool, which raises the importance of training. Users who understand how to describe an outcome, supply constraints, and review output will get much better results than those who treat it like search.

What to do: Review which users are in Frontier, then run a short prompting workshop for that group covering outcome-based requests, constraints, and review habits. Decide in advance whether the App skill is allowed to publish apps beyond the creator's team.

Agent2Agent and Federated Connectors Expand the Agent Mesh

Message center item MC1478975 announces that declarative agents in Microsoft 365 Copilot can now connect to Agent2Agent (A2A)-enabled Microsoft Agent 365 agents, alongside the existing OpenAPI and Model Context Protocol (MCP) options. The rollout began in late August and is expected to finish in late September worldwide. Separately, the September 23 release notes list general availability of federated connectors, which connect third-party data sources securely via MCP, along with an agent chat in Planner, expanded triage actions in Outlook, and a Copilot chat pane in Forms.

Microsoft rates the admin impact of the A2A change as high. Admins need to review and approve A2A-enabled agents on the All agents page of the Microsoft 365 admin center before users can reach them, and the Connected agents tab in a declarative agent's details shows what it is linked to. Agents calling other agents means data can flow in paths that no one designed explicitly, so approval processes and data-compliance reviews need updating.

What to do: Open the All agents page this week, inventory what is already approved, and add a step to your approval checklist that asks which other agents or MCP servers a new agent is allowed to call.

Copilot Studio: Credits, Governance, and a Billing Surprise

The most consequential Copilot Studio item from the September releases is financial. Agents built on the GitHub Copilot harness, now generally available, consume usage-based Copilot Credits regardless of Microsoft 365 Copilot licensing, and credits are spent while makers develop and test, not only when agents run in production. New controls arrived alongside it: environment-level enforcement rules, per-agent spending limits, and tenant-wide credit dashboards. Other September additions include native knowledge connections to Dataverse, Azure SQL, and SharePoint lists, an Agent Review tool for pre-release validation, and the ability to invoke agents as workflow steps with human approval gates before tool execution. Microsoft's documentation page for Copilot Studio had not yet listed September or October entries when we checked, so confirm details against the released-versions page.

This continues the dual-billing model discussed in recent weeks: a flat per-user license (about $30 per user per month for Microsoft 365 Copilot) for everyday AI, and metered credits for agentic work. Published estimates put a credit at roughly $0.008 to $0.01, and one agent response can consume anywhere from a handful to over a hundred credits, so costs are hard to predict without modeling.

What to do: Set tenant, group, and per-agent credit budgets with alerts before makers start building on the GitHub harness, and tell makers that testing consumes real credits.

Beyond Microsoft: Computer Use Comes to GitHub Copilot

On October 1, GitHub announced that GitHub Copilot can now interact with desktop applications in public preview, available in the Copilot CLI and the Copilot app on macOS and Windows. It can read app content, click controls, enter text, press keys, scroll, and drag, which opens automation to legacy and GUI-only software that has no API. Users see approval prompts before Copilot controls an app, can review or reset apps set to "always allow," and organizations can disable the feature through managed settings. The CLI toggle is /computer on.

This follows the same pattern as the computer-using agents in Copilot Studio: the interface becomes the API. It is powerful for old line-of-business tools, and it is also a new category of risk, because an agent with screen control inherits whatever the signed-in user can see and do.

What to do: Decide your policy now. If you do not want computer use on managed developer machines yet, disable it through managed settings, and document which legacy applications are good pilot candidates once you are ready.

Sources

Back to all news