Microsoft launches new Copilot with Home, Code and Autopilot for persistent AI work
Microsoft has introduced a major redesign of Microsoft Copilot, organizing the product around three new capabilities — Home, Code and Autopilot — as the company moves from conversational assistance toward persistent agents, delegated work and user-built software.
Home combines Chat and Cowork into a single starting point and brings Word, Excel and PowerPoint directly into the Copilot experience. Code lets users create applications, dashboards, automations and workflows from natural-language instructions using technology derived from GitHub Copilot. Autopilot goes further: it is a persistent cloud-hosted agent with its own identity, memory, computer and workspace, designed to continue working without waiting for another prompt.
Microsoft is also introducing Copilot Managed Runtime, an IT-governed environment for safely running the applications and automations produced by Copilot, while expanding Microsoft IQ, enterprise plugins and FinOps controls around long-running AI work.
Home and Code are beginning their rollout through Microsoft's Frontier program, with broader availability following in stages. Autopilot expands to private preview at the end of September.
........
NEW COPILOT | PRIMARY ROLE | EXECUTION MODEL |
Home | Chat, delegated work and Office documents | Interactive + delegated |
Code | Apps, dashboards, automations and workflows | User-created software |
Autopilot | Persistent recurring and proactive work | Long-running autonomous agent |
Copilot Managed Runtime | Secure execution environment | IT-governed infrastructure |
Microsoft IQ | Organizational context and grounding | Shared enterprise intelligence |
........
··········
HOME COMBINES CHAT, COWORK AND OFFICE INTO ONE WORKSPACE
Home becomes the central interface of the redesigned Copilot.
Microsoft is bringing Chat and Cowork together rather than presenting conversational assistance and delegated work as disconnected experiences.
Chat remains oriented toward immediate interactions such as questions, drafting, summaries, lookups and analysis.
Cowork is intended for tasks that can be delegated end-to-end. Instead of asking Copilot for one intermediate output, users can assign work such as preparing an RFP response, producing a customer briefing, assembling a launch package or creating material for a financial close.
Microsoft says a future routing layer will make the distinction increasingly automatic: users describe the objective and Copilot determines whether the task should be handled through Chat, Cowork or Code.
The practical change is that Copilot begins operating on a hierarchy of work rather than on one interaction format.
A quick question may require only inference. A complex deliverable may require prolonged document manipulation. A repeated business process may require software or an autonomous agent.
The same Copilot interface is being designed to route among those execution modes.
··········
WORD, EXCEL AND POWERPOINT NOW OPERATE INSIDE COPILOT
Microsoft is also integrating Word, Excel and PowerPoint directly into Copilot through Office in Copilot.
The objective is to reduce the handoff between asking the model to produce something and subsequently rebuilding that output inside an Office application.
A user can ask Copilot to draft a brief, construct a financial model, prepare a presentation or modify an existing team file, with the result created as an actual editable Office document rather than as text that must later be transferred manually.
The files remain collaborative. Users can edit the document beside the Copilot conversation, while colleagues can be mentioned and contribute changes that remain synchronized with the corresponding Office application.
This becomes particularly relevant for longer agent workflows because the deliverable is no longer separate from the system executing the work.
An agent producing a financial model, for example, can operate against the workbook that will ultimately be reviewed and used rather than generating a textual approximation of that workbook in chat.
··········
CODE LETS KNOWLEDGE WORKERS BUILD SOFTWARE IN NATURAL LANGUAGE
The second major component is Code.
Microsoft is positioning Code as a way for users outside traditional software-development roles to create small applications and automations directly inside Copilot.
A user can describe an app, tracker, dashboard, workflow or automation in natural language, and Copilot can determine an implementation approach and build it.
Microsoft gives examples ranging from persistent desktop widgets and interactive data dashboards to internally hosted applications that can be shared with a team.
Code uses the same underlying technology as GitHub Copilot, but its target audience and execution environment differ.
GitHub Copilot remains oriented toward professional software development. Code brings parts of the same agentic software-building capability into Microsoft Copilot for employees whose objective is solving a business problem rather than maintaining a conventional development repository.
........
EXAMPLE REQUEST | POSSIBLE CODE OUTPUT |
Monitor a recurring KPI | Persistent widget |
Explore business data | Interactive dashboard |
Automate an internal approval | Workflow |
Build a team utility | Internal application |
Connect operational data | Data-backed tool |
........
Microsoft's broader assumption is that the traditional document may no longer be the only basic unit of knowledge work.
Documents, spreadsheets and presentations remain central, but small pieces of purpose-built software can increasingly be generated for individual business processes.
That reduces the economic threshold for creating internal applications: a workflow that would previously have been too small to justify a dedicated development project may become viable when an agent can produce and maintain much of it.
··········
COPILOT MANAGED RUNTIME PROVIDES THE EXECUTION LAYER
Generating software is only one part of the problem.
Once users can create applications and automations freely, organizations also need somewhere to execute them with appropriate permissions, governance and security.
Microsoft is therefore introducing Copilot Managed Runtime.
The runtime provides hosting infrastructure for software created through Copilot and allows applications to execute inside the organization's Microsoft 365 environment while remaining governed by IT.
Applications can connect to live enterprise data, be shared with colleagues and remain available without requiring every user to provision independent infrastructure.
The same runtime is intended to support software generated through Cowork, Code and Copilot Studio, and Microsoft plans to expose the environment to professional and third-party developers as well.
Copilot Managed Runtime is currently in preview and will also be available from within Code.
This layer is important because widespread natural-language software creation creates a governance problem almost immediately.
Without centralized execution controls, organizations could end up with large numbers of employee-generated applications running against corporate systems with inconsistent permissions, logging and lifecycle management.
Managed Runtime gives Microsoft a way to pair easier software creation with centralized operational control.
··········
AUTOPILOT IS DESIGNED TO KEEP WORKING WITHOUT AN ACTIVE USER SESSION
Autopilot is the most substantial step away from the traditional assistant model.
Previously known as Scout, Autopilot is a persistent, proactive and personal agent that can continue working after the user leaves the conversation.
Users assign it a name, role and goal.
From there, the agent can monitor communication channels, follow up on discussions, execute recurring work and resume projects after long periods without requiring the user to initiate every step manually.
Microsoft describes scenarios such as managing a supplier-review process from start to finish: constructing the schedule, preparing meetings, following up with stakeholders and maintaining the work as new information arrives.
Because Autopilot is cloud-hosted, the process does not depend on the user's PC remaining active. The agent can continue operating while the user is offline.
··········
AUTOPILOT HAS ITS OWN IDENTITY, MEMORY, COMPUTER AND WORKSPACE
Microsoft describes each Autopilot agent as having four persistent resources: identity, memory, computer and workspace.
That architecture changes what persistence means.
A conventional assistant receives the current conversation plus whatever data the application retrieves for the request. A persistent agent needs continuity across much longer periods.
Its identity determines where it can appear and which permissions apply. Its memory allows previous work and decisions to influence future actions. Its computer gives it an execution environment. Its workspace provides persistent state for the project or responsibility it has been assigned.
Autopilot lives inside the organization's tenant and can appear across Teams, Outlook, chats, channels and documents.
Users can mention it similarly to a colleague, while enterprise permissions, auditability and governance remain attached to its activity.
This creates a different organizational object from a chatbot session.
The agent can increasingly resemble a persistent service account attached to a role and objective, but with a conversational interface and an execution environment behind it.
··········
MICROSOFT IQ PROVIDES THE BUSINESS CONTEXT BEHIND THE AGENTS
Home, Code and Autopilot rely heavily on Microsoft IQ, Microsoft's intelligence layer for grounding agents in organizational data and business context.
Microsoft is expanding the information available through that system.
Fabric IQ
Fabric IQ can provide trusted business context derived from enterprise data, including semantic models built in Power BI. That information is already available to Chat and Cowork, while Code integration is planned through Frontier.
Microsoft is also adding context from Dynamics 365 and Power Platform, allowing Copilot to combine document-level work with information held inside operational business systems.
For example, an agent preparing a commercial proposal could incorporate customer history, support information and deal data rather than requiring the employee to manually assemble those inputs before asking Copilot to work.
The quality of persistent agents depends heavily on this layer because autonomy without current business state quickly becomes unreliable.
Autopilot needs to know not only what the user requested originally, but also what has changed in the organization since the task began.
··········
A UNIFIED PLUGIN REGISTRY EXPANDS WHAT COPILOT CAN ACTUALLY DO
Microsoft is also consolidating extensibility around a new plugin registry.
The registry brings Microsoft-built, partner-built and organization-specific plugins into one catalog.
Plugins combine capabilities such as connectors, skills and actions, allowing Copilot and its agents to interact with external systems rather than remaining limited to information already available inside the model context.
IT administrators can approve and manage plugins centrally, while developers can publish a capability once and expose it across supported Copilot experiences.
This becomes increasingly important as Copilot moves from answering questions toward executing processes.
A model can reason about a supplier review using retrieved documents, but completing the process may require updating systems, creating records, contacting participants and changing workflow state.
The plugin layer provides the connection between reasoning and those external actions.
··········
AUTOPILOT CHANGES THE COST MODEL OF COPILOT
Persistent agent work also creates a different economic structure from ordinary AI assistance.
Microsoft is separating conventional user interactions from more compute-intensive agent workloads.
A user subscription license covers everyday tasks such as drafting, summarization, analysis and conversational Copilot use.
For those interactions, Microsoft's Auto system can route requests among models based on a balance of accuracy, speed and cost.
Long-running work uses usage-based billing.
Cowork, Code and Autopilot, along with frontier models and other extended agent workloads, can therefore consume metered resources based on actual execution.
........
WORKLOAD | PRIMARY BILLING MODEL |
Chat and everyday Copilot | User subscription license |
Standard Office assistance | User subscription license |
Cowork tasks | Usage-based billing |
Code execution | Usage-based billing |
Autopilot | Usage-based billing |
Frontier-model agent work | Usage-based billing |
........
The distinction is economically necessary because a five-second question and an autonomous process running intermittently for several days cannot reasonably have the same marginal cost structure.
For persistent agents, organizations must measure more than token prices.
A useful operating metric becomes cost per successfully completed business process, incorporating model inference, tool calls, runtime usage, failures, retries and human intervention.
··········
MICROSOFT IS ADDING FINOPS CONTROLS SPECIFICALLY FOR AI AGENTS
Microsoft is expanding its FinOps for AI tooling alongside the new Copilot architecture.
Agent 365 cost management is being extended to Code and Copilot Managed Runtime, with additional agent environments planned later.
Administrators can create spending policies, manage credit requests, determine which model families are available to different employee groups and use those restrictions to influence which models automatic routing can select.
Users can also see their own credit consumption, remaining balance and usage history directly inside Copilot.
For business leaders, Microsoft is adding reporting intended to connect agent consumption to outcomes.
That becomes necessary once long-running agents begin producing variable costs independently of direct employee interaction.
A company operating hundreds or thousands of agents needs to understand which automated processes are economically productive rather than merely knowing how much inference was consumed.
··········
TODAY WILL MAKE HOME MORE PROACTIVE
Microsoft has also previewed Today, a forthcoming component of Home.
Today is designed as a personalized command center spanning email, calendars, Teams discussions, meetings and tasks.
Instead of only presenting information, it can prepare work before the user explicitly asks for each individual action.
Examples include drafting a response that is waiting for review, preparing a proposed schedule change or surfacing an overlooked follow-up with the relevant context already attached.
Today is scheduled to enter private preview in October and is also planned for Outlook and Teams.
This introduces proactive behavior at a different level from Autopilot.
Autopilot owns persistent delegated responsibilities. Today focuses on continuously preparing the employee's immediate work environment.
The two approaches could eventually converge around a system in which some tasks remain human-driven, some arrive pre-prepared for approval and others are delegated completely to persistent agents.
··········
@COPILOT IN TEAMS WILL OPERATE WITH SHARED GROUP CONTEXT
Microsoft is also preparing a new @Copilot experience in Teams.
When invoked inside a channel, meeting or group chat, Copilot can use the context and permissions associated with that shared workspace rather than operating only from one individual's private context.
This allows teams to ask questions about previous decisions, project dependencies or unresolved issues using information accumulated across collaborative discussions.
Microsoft says Copilot can then move beyond retrieving the explanation and offer to prepare follow-up work such as a cross-functional review containing dependencies, owners, unresolved questions and decisions.
The capability is expected to enter private preview by the end of September.
The permission model is significant because collaborative agents need to distinguish between information available to an individual employee and information legitimately shared within the group where the agent is operating.
··········
HOME, CODE AND AUTOPILOT CREATE THREE DIFFERENT LEVELS OF AI WORK
The redesigned Copilot effectively introduces a progression from assistance to delegation to persistent execution.
........
LEVEL | USER ROLE | COPILOT ROLE |
Chat | Drives interaction | Responds |
Cowork | Defines task and reviews result | Executes delegated work |
Code | Describes required solution | Builds software |
Autopilot | Sets role, goal and boundaries | Operates persistently |
........
The distinction becomes more important as the duration of AI work increases.
A chat exchange might last seconds. A Cowork task may run long enough to produce a complex deliverable. A Code application may continue operating indefinitely after it has been generated. An Autopilot assignment can remain active for days or longer, reacting to new events without requiring another initiating prompt.
The technical requirements expand accordingly: persistent state, identity, permissions, runtime infrastructure, auditing, cost controls and organizational context become increasingly important as execution becomes less dependent on active user supervision.
··········
THE NEW COPILOT MOVES MICROSOFT FROM ASSISTANT SOFTWARE TOWARD AN AGENT OPERATING LAYER
Home, Code and Autopilot are not three versions of the same chatbot.
They represent separate execution modes inside an increasingly unified environment for knowledge work.
Home organizes interaction and delegation. Code converts natural-language requirements into executable software. Managed Runtime provides infrastructure for that software. Microsoft IQ supplies organizational context. Plugins connect agents to external actions. Autopilot supplies persistent identity, memory and execution.
For organizations adopting the system, this shifts the performance question away from how well Copilot responds to an individual prompt.
The more relevant measurements become how much work reaches completion, how often humans must intervene, how reliably agents maintain state across long processes, what permissions they exercise, and how much each completed workflow costs.
Microsoft's new Copilot architecture is being built around those longer execution cycles.
The result is a product that increasingly treats conversation as only one interface into a broader system of delegated tasks, generated software and persistent AI workers.
·····
FOLLOW US FOR MORE.
·····
·····
DATA STUDIOS
·····
[datastudios.org]
