Kommentar
Åtkomst till den här sidan kräver auktorisering. Du kan prova att logga in eller ändra kataloger.
Åtkomst till den här sidan kräver auktorisering. Du kan prova att ändra kataloger.
This article describes the fixes and improvements to existing functionality that are included in this build.
Visit the Microsoft Copilot Studio Blog to stay up-to-date on the latest news!
Version 2026.5.3 updates
New features
Environment-wide analytics overview across multiple agents provides total sessions, autonomous runs, enhanced task completion (preview) sessions, daily active user averages, and day-by-day trends for a selected date range.
Custom metrics let makers define, generate, and track metrics for agents in Analytics, with creation less likely to be blocked by content moderation while appropriate filtering still applies.
Maker Evaluation supports running general quality evaluations on workflow runs and handles invalid workflow inputs cleanly without triggering evaluation calls.
Model Context Protocol tooling allows listing and previewing tools exposed by an MCP-enabled Power Platform connector (for a selected connection) before adding the connector to an agent.
Speech and voice updates include automatic speech recognition domain language model support enabled by default for prebuilt and closed list entities, and default voice selection that respects an agent’s configured voice processing mode with environment-level lists defaulting to Classic mode.
Breaking change: custom pronunciation lexicons are limited to 500 entries per lexicon entity in private preview; training with more than 500 entries fails with an actionable error while other training continues.
Fixes and improvements
Continued improvements to performance and reliability.
Enhanced task completion (preview): agents that reference flows publish successfully; external trigger callbacks route to the agentic runtime; evaluation runs execute against the agentic runtime endpoint; and missing flow trigger callback URLs no longer cause runtime failures after publish.
Evaluations and scoring: maker evaluation retrieves transcripts for enhanced task completion (preview) agents; OpenAI request timeouts fail in a predictable, controlled way; maker evaluation scoring and custom metric generation are less likely to be blocked by responsible AI filtering; and analytics uses the latest published agent version when processing older summaries.
Connectors and tools: connector actions prompt for reconnection when connections are stale or deleted; Microsoft 365 tool calls are less likely to time out; connector discovery no longer returns “not found”; Copilot plugin operations tolerate invalid or missing parameter type information and log warnings; agents reuse user-owned connections more consistently; tool/function input schemas are accepted more consistently by provider APIs; MCP manifests include required authentication configuration and allowed tools; tools from multiple MCP servers with the same name no longer conflict; and third‑party sign‑in loops (Generic OAuth 2.0) are resolved.
Knowledge sources and citations: SharePoint setup more reliably discovers document libraries (including across multiple pages); search-based responses stream more reliably with consistent citation numbering between streamed and final responses; repeated references use consistent citation numbers; and SharePoint knowledge sources uploaded to agents are correctly tracked and enforced by Data Loss Prevention and modern governance.
Authoring, runtime, and UX: PDFs are processed and summarized correctly (including fewer-page PDFs); non-text file uploads on the first message no longer cause agent failures and can be used downstream; skill file attachments from Dataverse load with correct content and type; live chat widget and digital test chat route and stream responses more reliably; escalation to human includes the agent message and additional context; connected agent access errors no longer stall the main agent; analytics summary pages show labels from the latest draft sooner; session analytics for enhanced task completion (preview) agents include transcript messages; analytics themes classification is more reliable; environment-level voice list retrieval no longer returns 404; multi-digit keypad input is interpreted more reliably; only workflows referenced by an agent are returned when requested; primary owner is reassigned automatically when removed; and filtering agent activity limits results to a recent time window (default last 30 days).
Security
Messages marked as sensitive are no longer forwarded in conversation history when invoking other external agents.
Data Loss Prevention and modern governance policies are enforced for SharePoint knowledge sources, including federated SharePoint sites, which might block answers or raise errors when content is restricted.
Cross‑cloud authentication supports organizations using Microsoft Copilot Studio in Azure US Government to publish agent inventory to Azure Public more reliably.
Validation for adding a skill whose app registration or agent identity is not in the current tenant returns a clear “skill must be registered” message instead of a server error.
Return to the Microsoft Copilot Studio Release Notes.