Copilot Studio in 2026: Low-Code Agents Reach Production
Copilot Studio entered 2026 with four production-relevant capabilities: model choice including Anthropic's Claude models, computer-use automation in preview, generally available MCP integration, and agent evaluations. We examine what each capability shipped, what remains preview-only, and why low-code agents now clear the bar for production use — with dates, limits, and open questions.
The low-code credibility problem
Low-code agent platforms have carried the same reputation for years: convincing in the demo, fragile in production. Microsoft Copilot Studio, launched in November 2023 as the successor to Power Virtual Agents, was no exception. Until early 2025 it offered one model family, no standard protocol for connecting tools, and no automated testing. Teams that needed reliability built their agents in code.
That changed within twelve months. Four capabilities arrived in sequence: model choice including Anthropic models, computer-use automation, Model Context Protocol (MCP) integration, and agent evaluations. As of February 2026, the four together justify a reassessment. This article states what each capability does, what it does not do, and what remains open.
Multi-model choice with Anthropic models
On September 24, 2025, Anthropic models joined OpenAI models in Copilot Studio: Claude Sonnet 4 and Claude Opus 4.1, selectable for orchestration, chat, and deep-reasoning scenarios. Claude Sonnet 4.5 followed on September 29 — the day of its public release. On November 24, Claude Opus 4.5 arrived as an experimental model, again on release day. Microsoft stated in November that GPT-5, GPT-5.1, Claude Sonnet 4.5, and Claude Opus 4.5 had each been enabled on the day of their public release. GPT-5 Chat reached general availability for US and European regions on November 24, 2025.
The scoping matters. OpenAI remains the default for new agents. Anthropic models require an administrator opt-in in the Microsoft 365 Admin Center, are hosted outside Microsoft-managed environments, and are subject to Anthropic's terms of service. If an administrator disables them, agents fall back automatically to the default model. Depending on region, requests may be processed cross-geo. Model choice in Copilot Studio is therefore a governance decision as much as a quality decision.
Computer use agents remain in preview
Computer use was announced on April 15, 2025 as an early-access research preview: agents operate websites and desktop applications through the graphical interface — clicking, typing, navigating — where no API exists. The public preview followed on September 15, 2025 for US-based environments, adding a hosted browser on Windows 365, a credential store, and allow-lists that terminate a run when the agent leaves approved applications. January 2026 brought audit logging with session replay and Cloud PC pooling, both in preview.
What computer use does not offer: general availability. As of this publication it remains a preview feature; Microsoft's release plan targets GA for May 2026. Vision-based automation is also slower and more expensive per step than a direct API call. Our working rule is unchanged: API first, MCP second, computer use only where neither exists.
MCP makes integration a standard
Copilot Studio shipped MCP support in preview on March 19, 2025 and declared it generally available at Build on May 19, 2025 — the first of the four capabilities to reach GA. Tools published by an MCP server appear automatically in the agent and track server-side changes without redeployment. At Ignite in November 2025, Microsoft counted more than 1,400 connectable systems across MCP, Power Platform connectors, and Microsoft Graph.
MCP connections run through the existing connector infrastructure, which brings virtual network integration, data-loss-prevention policies, and multiple authentication methods. Generative orchestration is required. One caveat stands: MCP standardizes transport, not trust. Tool names and descriptions from a server enter the agent's context and must be reviewed like third-party code.
Agent evaluations add measurement
Agent evaluations entered public preview on October 27, 2025. Makers build test sets from uploaded files, recent Test Pane interactions, manual cases, or AI-generated queries derived from the agent's metadata and knowledge sources. Graders range from exact and partial match through similarity and intent match to AI-scored relevance, completeness, and groundedness. December 2025 added side-by-side comparison of agent versions; January 2026 added activity maps for tracing a run and a feedback channel on grader quality.
The limits are documented. Evaluations are single-turn in the current preview; multi-turn testing is on the roadmap. AI graders are themselves models and can misjudge — the feedback channel exists precisely for that. An evaluation set does not guarantee correctness. It makes regressions visible before users find them. That is the point.
What production credibility looks like
Production credibility is a checklist, not a feeling: governed model choice, tested behavior, auditable actions, and lifecycle tooling. The last item closed most recently. Microsoft Entra Agent ID, in preview since November 2025, assigns agents their own identities. The Copilot Studio extension for Visual Studio Code reached general availability on January 14, 2026, making agent definitions diffable, reviewable, and versionable. We now treat Copilot Studio agents in customer projects as software artifacts — because the tooling finally permits it.
| Capability | First preview | Status on 12 Feb 2026 |
|---|---|---|
| Anthropic models | Sep 24 2025 | Rolled out; admin opt-in required |
| Computer use | Apr 15 2025 | Public preview (US environments) |
| MCP integration | Mar 19 2025 | GA since May 2025 |
| Agent evaluations | Oct 27 2025 | Public preview |
Outlook from February 2026
From today's vantage point we expect three developments. First, computer use reaches general availability in May 2026 as planned, and UI automation shifts from RPA licenses into agent platforms. Second, evaluations grow multi-turn and move into deployment pipelines — an agent version that fails its evaluation set will not ship, the same gate a failing unit test provides today. Third, model churn accelerates: frontier release cycles are now measured in weeks, and pinning an agent to a single model without an evaluation set is technical debt.
Some of this will prove wrong; forecasts written on publication day always do. The direction is harder to miss. The gap between low-code and pro-code agent development is closing from both sides, and platforms that combine model choice, standard protocols, and measurement will define the production baseline. We plan accordingly.
Sources
- Anthropic joins the multi-model lineup in Microsoft Copilot Studio (24 Sep 2025)
- Computer use is now in public preview in Microsoft Copilot Studio (15 Sep 2025)
- Model Context Protocol (MCP) is now generally available in Microsoft Copilot Studio (29 May 2025)
- Build smarter, test smarter: Agent Evaluation in Microsoft Copilot Studio (27 Oct 2025)
- What's new in Copilot Studio: November 2025 (15 Dec 2025)
- Copilot Studio Extension for Visual Studio Code Is Now Generally Available (14 Jan 2026)