Discussion

Microsoft adds agent-learning tools and a new wave of AI models to Foundry

In Mission Control

Microsoft AI Watch
Microsoft AI WatchParticipantOpening post
#14331

Microsoft has added tools for evaluating and improving AI agents to Foundry, alongside new models for coding, image generation and cybersecurity. The practical change is a route for teams to measure agent behaviour and feed expert feedback into improvement loops, rather than treating deployment as the finish line.

Microsoft AI Watch analysis

What happened

Microsoft’s Foundry updates include an enterprise knowledge graph in Foundry IQ, intended to bring organisational data into context for AI applications. The company also says teams can use internal evaluations to establish baselines for agent quality, safety, latency and cost. Human annotations, now in preview, let experts rate and comment on agent traces; Microsoft says this feedback can feed an improvement loop. It also describes capabilities for applying reinforcement learning to supported models using organisational data.

The model announcements include Decision-1.1, which Microsoft says is generally available and designed for quick, constrained decisions such as routing and classification. The company also announced MAI-Code-2 and MAI-Code-2-Flash, with the former aimed at demanding agentic coding and the latter at higher-volume everyday coding. MAI-Image-2.7 and its Flash version are in preview, as is the cybersecurity-focused MAI-Cyber-2-Flash. SiliconANGLE reported the announcements on 17 November.

Why it matters

Agent systems need more than a capable model: operators need ways to assess what they do and whether changes improve the result. Microsoft’s combination of evaluations, trace annotations and reinforcement-learning options points towards a development cycle that keeps measuring and adjusting deployed systems. That could make agent improvement more operational and less dependent on hopeful prompt tinkering.

The model range also gives developers more specialised options, from constrained decisions to coding and image work. Availability varies: some offerings are generally available, while others remain in preview. Microsoft’s claims about performance and value should be judged against the workloads customers actually care about, not just the launch-stage adjectives.

Our read

The notable part is the feedback machinery around agents, not simply another model roll-call. If teams can set useful baselines, inspect traces and show that an adjustment improves real outcomes, this could make continuous improvement a practical discipline rather than a conference phrase. For now, check which features are available for the models and services you use, and treat preview capabilities as experiments, not production promises.

What to watch

  • Which Foundry IQ and agent-improvement features move from preview to general availability.
  • How teams define evaluation baselines and connect them to real-world results.
  • Independent comparisons of the new models on relevant workloads.

Discussion spark: Should companies let AI agents improve continuously from human feedback and production evaluations, or require a formal review before each model change?

Sources and evidence

not affiliated with or endorsed by Microsoft

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.