AI

Microsoft ships Decision-1.1 model and continuous learning for AI agents

Foundry platform gains human-in-the-loop feedback, reinforcement learning, and expanded model catalog as Microsoft pushes enterprise agentic AI forward.

Omega Editorial · November 17, 2026 · 3 min read

Key takeaways

  • Microsoft's Decision-1.1 model handles high-volume software decisions faster and cheaper than general-purpose LLMs by working within constrained parameters and providing confidence scores.
  • Foundry now enables continuous improvement loops for AI agents through human annotations, reinforcement learning, and automated optimization tests against real-world results.
  • The MAI model family expanded with new coding models (MAI-Code-2), image generation models (MAI-Image-2.7), and a cybersecurity-focused model (MAI-Cyber-2-Flash).
  • Foundry's model catalog now includes more than 11,000 models from major providers, all accessible through unified evaluation and deployment workflows.
  • Human-in-the-loop feedback allows domain experts to rate agent performance based on business judgment, feeding directly into automated improvement cycles.

Microsoft advances enterprise AI with new decision model and agent improvement tools

Microsoft has launched Decision-1.1, a specialized AI model designed to handle high-volume decision-making tasks in enterprise software, alongside significant updates to its Foundry platform that enable autonomous agents to learn and improve continuously from real-world performance.

The Decision-1.1 model, now generally available, addresses a specific inefficiency in how large language models handle routine software decisions. Rather than using computationally expensive LLMs to "think through" tasks like routing, classification, prioritization, and verification, Decision-1.1 reaches answers quickly within constrained parameters and provides calibrated confidence scores. When uncertainty is high, applications can defer to human judgment or more capable models.

Built on Google's Gemma 4 26B A4 base, the new model adds multimodal support for analyzing images in instructions while delivering improvements in both intelligence and speed compared to the October release of Microsoft-Decision-1, according to details first reported by SiliconANGLE.

Continuous improvement loops for autonomous agents

Foundry, Microsoft's unified enterprise AI platform, now incorporates transparency and observability capabilities into learning loops that allow autonomous agents to improve over time. Internal evaluations establish baselines for quality, safety, latency, and cost, creating a foundation for behavioral improvements.

A key addition is human annotations, currently in preview, which let domain experts rate and comment on agent traces based on business judgment. This feedback feeds directly into optimization tests that compare results against evaluations and real-world outcomes. Foundry can now adapt supported models with organizational data and apply reinforcement learning during runtime, with each improvement cycle becoming the next baseline for measurement.

Foundry IQ updates deliver an enterprise knowledge graph out of the box, providing access to unstructured and application data across an organization's full estate to support richer context for AI applications.

Expanded model family and catalog

Microsoft extended its MAI first-party model family with several new releases. MAI-Code-2 handles demanding agentic coding work, while MAI-Code-2-Flash supports high-volume everyday coding at lower cost. Both are generally available on GitHub, with Code-2-Flash in private preview within Foundry.

For visual content, MAI-Image-2.7 and MAI-Image-2.7-Flash, now in preview, offer sharper editing and more consistent image generation to help brands maintain visual consistency across campaigns. In cybersecurity, MAI-Cyber-2-Flash is purpose-built for security efforts and available in preview within Codename MDASH, Microsoft's agentic code security system.

The Foundry platform now features an expanded model pool with the top 50 models, supporting a model router and AI Gateway for intelligent model selection. The catalog includes more than 11,000 models from OpenAI, Anthropic, Meta, Mistral AI, and the open-source community, all of which can be evaluated and deployed using the same continuous improvement workflow applied to agents.

Why it matters

The combination of specialized decision models and continuous learning infrastructure addresses two critical enterprise AI challenges: reducing the cost and latency of routine AI tasks while creating systematic pathways for agents to improve from production experience. Human-in-the-loop feedback mechanisms bridge the gap between automated evaluation and business judgment, potentially accelerating the path from prototype to production-grade autonomous systems. For organizations deploying AI agents at scale, the ability to measure, test, and optimize behavior continuously could determine whether agentic AI delivers sustained value or remains a high-maintenance experiment.

SiliconANGLE first reported these updates on November 17.

  • #microsoft
  • #agentic ai
  • #decision models
  • #continuous learning
  • #foundry platform
  • #enterprise ai

This is an original analysis by the Omega editorial team. Source reporting: AI Watch.

Want systems like this working for your business?

Book a Call

More in AI