In September 2026, Microsoft shipped another significant update to Microsoft Foundry, its platform for building and running AI agents.
What was announced
- Routines, generally available: run agents automatically on a schedule, after a delay or in response to an event, without building a separate scheduler.
- Voice agents, public preview: a new agent type in Foundry Agent Service for both prompt and hosted agents, with voice-agent observability included.
- Tool search in Toolboxes, generally available: agents can find the tools they need without loading the whole tool catalogue upfront. Toolboxes are generally available for hosted agents and in preview for prompt agents.
- Resilience for long-running hosted agents, public preview: work can continue after a client disconnects and recover if the hosting process stops unexpectedly.
- Optimisation: Insights in Foundry is in public preview, with the rubric evaluator, traces-to-dataset generation and agent optimizer announced for general availability later that month.
- Models: Microsoft described the GPT-6 family from OpenAI and Claude Opus 5.5 from Anthropic as available in Foundry.
What this unlocks
- Scheduled back-office work: reconciliations, report preparation and queue triage can run overnight on routines.
- Voice in customer service: voice agents give organisations a path to natural-language phone experiences on the same platform as their other agents. That is relevant across African markets, where voice remains a primary channel.
- Reliability: long-running tasks that survive disconnects suit environments with less predictable connectivity.
Practical advice
- Choose models by task, not hype. Larger models cost more. Evaluate quality, latency and cost for each workload.
- Use evaluations before and after launch. The rubric evaluator and tracing make quality measurable.
- Keep humans in the loop for decisions with financial or customer impact.
Planning an agent on Azure? Our AI & Copilot team can help you design, evaluate and govern it.



