Microsoft-Decision-1: Decision model for controlling AI agents
Microsoft introduces a specialized model for routing, classification and agent control. Performance claims are vendor-reported.
Microsoft-Decision-1 is a model for scoring predefined choices, including agent handoffs, model routing and answer evaluation. It is not a standalone AI agent, and the reported performance figures come from Microsoft’s own evaluations.
On October 9, 2026, Microsoft introduced Microsoft-Decision-1 as a specialized model for structured decision tasks. According to the product announcement, it takes a fixed set of answer options and returns a probability score for each. Microsoft highlights classification, prioritization, validation and workflow control as intended applications. Access is described through Microsoft Foundry and OpenRouter. This does not establish general availability in every Copilot surface or agent product.
A decision model is not the same thing as an acting agent. Decision-1 does not, by itself, complete a research project, schedule persistent tasks or execute arbitrary third-party tools. An existing orchestrator can instead use its score to select a next action: continue, stop, retry, delegate to another system or ask for human review. The surrounding workflow still needs to enforce identity, permissions, error handling and escalation rules.
Microsoft describes the model as post-trained from Qwen3.5-9B and reports comparisons across 36 benchmarks and nearly 150,000 questions. Claims about speed, consistency and cost also come from Microsoft’s own testing. AgentenCode does not present these as independently replicated findings or as system-level performance guarantees. Comparisons with different orchestration patterns would require aligned datasets, versions, execution environments and evaluation criteria.
The operational distinction is between a decision score and the authority to perform an action. Confidence values are not proof that a proposed action is safe, compliant or appropriate for an organization. Sensitive workflows still require enforceable policies, recorded decisions and, where appropriate, human approval. This announcement is relevant to orchestration and evaluation nodes; it does not create a new AgentenCode agent profile or automatically satisfy any AgentenTrust control.
Primary sources
Original sources and exact scope
External product documentation; no independently measured system-wide advantage or implied capabilities.