Microsoft published its October 2026 Microsoft Foundry update on 5 November, with the Copilot harness entering public preview and long-running hosted-agent resilience reaching general availability. The same update brings Browser Automation and the Decisions API to general availability, giving development teams more managed options for agent execution, browser work and typed routing decisions.
The common thread is task completion. Microsoft is moving more of the agent loop, browser session and recovery handling into Foundry services, while leaving application teams responsible for checkpoints, approvals, permissions and evaluation.
Copilot harness preview
The Microsoft Copilot harness in Foundry Agent Service was announced on 28 October. It gives prompt agents a managed loop and sandbox for tasks that require file handling, code execution, supported tools and reusable skills, rather than a single model response.
Developers still choose a supported model. The harness manages how the model’s next step becomes work inside the configured tools, permissions and policy limits. Agent definitions can be source-controlled, and reusable skills can carry repeatable procedures alongside application code.
The preview has an important deployment constraint. Teams must select the harness when creating the agent, because switching an existing agent to the harness is not yet supported. Existing agents are unchanged, and Foundry manages the sandbox CPU, memory and idle settings.
Billing for the Copilot harness in Foundry Agent Service starts on 1 December 2026. Harness runtime usage is free before that date, but model inference and separately billed tools or connected services still incur their applicable charges. From 1 December, Microsoft lists consumption pricing at USD 0.0994 per vCPU-hour and USD 0.0118 per GiB-hour for sandbox compute and memory, with prices varying by region and agreement.
Recovery is GA, but checkpointing remains application work
Long-running resilience for hosted agents also reached general availability on 28 October. Foundry can restart work with its original input after an interruption, which matters for agents running tasks that take hours or days.
The service does not remove the need for application-level recovery design. Developers still need to save intermediate results, completed steps and identifiers that prevent repeated external actions. A restart after a report has been saved, for example, should reuse the report and check whether a review request already exists before sending another one.
This makes interruption testing part of the production readiness work. Teams need to test recovery around each boundary where the agent writes to another system, including the gap between an external service accepting a request and the application recording that acceptance.
Browser Automation and typed decisions reach GA
Browser Automation became generally available on 19 October. It uses a separately provisioned Playwright Workspace and connects through a Toolbox remote Model Context Protocol endpoint. Microsoft says it supports prompt and hosted agents, public and private websites, and private-network connectivity configured with the workspace.
The feature is intended for tasks that require a web interface, such as checking a legacy application status page or stepping through an internal web process. Microsoft’s guidance keeps the browser identity least-privileged and calls for human approval before consequential actions such as submitting a transaction or changing access. Sensitive-content protection for traces remains in preview.
The Decisions API in Foundry Models reached general availability on 22 October. It lets an application ask a typed question about shared input and receive a typed answer, rather than parsing generated text. The supported question types are predicate, choice and score.
This fits routing and bounded evaluation use cases. A predicate can estimate whether an observable condition is true, a choice can select from a fixed set, and a score can place an item on an ordered rubric. The API supports gpt-6-luna, with East US 2 excluded from the rollout at the time of the update.
Model and local runtime changes
GPT-6.1 Terra became generally available in Foundry on 26 October. It accepts text and images through the Responses API for visual reasoning and screen-based agent tasks. Microsoft lists support for up to one million input tokens, with Global Standard and Data Zone Standard deployments.
The October update also adds model catalogue movement outside the main agent runtime. Model Router version 2025-11-18 supports 35 underlying models after eight October additions. MAI Image 2.6 is in preview in East US, West Central US, West US, Sweden Central, West Europe and South India.
For speech and audio, Microsoft introduced MAI-Transcribe-2-Streaming, MAI-Voice-2.1 and MAI-Voice-2.1-Flash on 1 October. MAI-Transcribe-2-Streaming supports 60 languages, automatic continuous language detection, and partial and final transcripts. Introductory pricing is USD 0.54 per audio hour through the end of 2026, while MAI-Voice-2.1 is listed at USD 22 per 1M characters and MAI-Voice-2.1-Flash at USD 15 per 1M characters.
Foundry Local 2.2.1 shipped on 23 October with coding-model support, tool-calling changes, catalogue v2, corrupted-cache removal and reliability fixes. A separate CLI 0.12.1 shipped in public preview on the same day. GPU applications need particular care, because CUDA and cuDNN runtimes must be installed separately and the required CUDA major differs between x64 and ARM64.
SDK upgrades need testing
The October tooling updates include useful discovery work, but also migration risk. Foundry Toolkit 1.6.18 for Visual Studio Code adds a redesigned hosted-agent sample gallery, preview Copilot harness support in Agent Builder, and model filters for task, capability, region, deployment option, SKU and lifecycle.
The SDK changes are not uniform across languages. Python 2.8.0 and 2.9.0 include optimisation, data generation and evaluator lifecycle APIs, preview harness support and preview voice model renames. JavaScript and TypeScript 2.8.0 and 2.9.0 move some operations from beta groups while other scenarios still need preview headers. The .NET Projects package remains beta, and the Java SDK changes include direct dataset and evaluator clients replacing beta clients.
Azure Developer CLI 1.37.2 shipped on 28 October. Its agents extension remains beta and reached 1.0.0-beta.23 on 29 October, including a --harness option on azd ai agent init. Microsoft’s practical guidance is to upgrade the application and its adapters as a tested set, with regression checks for tool connection and cleanup, approval handling, cancellation and recovery around saved checkpoints.
Managed agent loops, browser-driven tasks, typed decisions and interruption recovery are the production boundaries that ECS's Core AI topic deals with when it covers agentic AI, Azure AI Foundry, the EU AI Act and responsible AI. Sessions and practitioner experience give attendees a place to compare service features with the engineering controls still needed around identity, approvals, evaluation and cost. Bring the agent scenarios where the documentation leaves design choices open, and use the conference to test them with people doing this work in practice.
This article was written by the ECS editorial team with the assistance of Amal Agent, run.events' AI assistant, which supported the research, drafting and proofreading.